Update README.md

README.md

@@ -4,13 +4,11 @@ emoji: 🐨
 colorFrom: green
 colorTo: indigo
 sdk: gradio
-sdk_version: 5.44.1
+sdk_version: 5.48.0
 app_file: app.py
 pinned: false
 license: mit
-python_version: "3.13"
 short_description: AI-powered image editing tool
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
-
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

__lib__/app.py

@@ -1,1455 +0,0 @@
-import gradio as gr
-import threading
-import os
-import shutil
-import tempfile
-import time
-import json
-from util import process_image_edit, download_and_check_result_nsfw, GoodWebsiteUrl
-from nfsw import NSFWDetector
-
-# Google Gemini URL for restricted languages
-GOOGLE_GEMINI_URL = "https://aistudio.google.com/models/gemini-2-5-flash-image"
-
-# i18n - Load from encrypted modules
-import sys
-from pathlib import Path
-
-# Add i18n module to path
-_i18n_module_path = Path(__file__).parent / "i18n"
-if str(_i18n_module_path) not in sys.path:
-    sys.path.insert(0, str(_i18n_module_path))
-
-# Import encrypted i18n loader
-from i18n import translations as _translations
-translations = _translations
-
-def load_translations():
-    """Compatibility function - translations are already loaded"""
-    return translations
-
-def t(key, lang="en"):
-    return translations.get(lang, {}).get(key, key)
-
-# Configuration parameters
-# a = b
-TIP_TRY_N = 1      # Show like button tip after x tries
-FREE_TRY_N = 4        # Free phase: first 15 tries without restrictions
-SLOW_TRY_N = 6       # Slow phase start: 25 tries
-SLOW2_TRY_N = 10       # Slow phase start: 32 tries
-RATE_LIMIT_60 = 14     # Full restriction: blocked after 40 tries
-
-# Time window configuration (minutes)
-PHASE_1_WINDOW = 6     # 15-25 tries: 5 minutes
-PHASE_2_WINDOW = 13    # 25-32 tries: 10 minutes
-PHASE_3_WINDOW = 20    # 32-40 tries: 20 minutes
-MAX_IMAGES_PER_WINDOW = 2  # Max images per time window
-high_priority_n = 1 # 每个ip只有第一个任务是高优先级的
-
-IP_Dict = {}
-# IP generation statistics and time window tracking
-IP_Generation_Count = {}  # Record total generation count for each IP
-IP_Rate_Limit_Track = {}  # Record generation count and timestamp in current time window for each IP
-IP_Country_Cache = {}  # Cache IP country information to avoid repeated queries
-
-# Country usage statistics
-Country_Usage_Stats = {}  # Track usage count by country
-Total_Request_Count = 0  # Total request counter for periodic printing
-PRINT_STATS_INTERVAL = 10  # Print stats every N requests
-
-# Async IP query tracking
-IP_Query_Results = {}  # Track async query results
-# Active task tracking (within recent time window)
-Active_Tasks = {}  # {client_ip: {"start": timestamp}}
-
-# Restricted countries list (these countries have lower usage limits)
-RESTRICTED_COUNTRIES = ["印度", "巴基斯坦", "俄罗斯", "中国", "伊朗"]
-RESTRICTED_COUNTRY_LIMIT = 1  # Max usage for restricted countries
-
-country_dict = {
-    "zh": ["中国"],
-    "hi": ["印度"],
-    "fi": ["芬兰"],
-    "en": ["美国", "澳大利亚", "英国", "加拿大", "新西兰", "爱尔兰"],
-    "es": ["西班牙", "墨西哥", "阿根廷", "哥伦比亚", "智利", "秘鲁"],
-    "pt": ["葡萄牙", "巴西"],
-    "fr": ["法国", "摩纳哥"],
-    "de": ["德国", "奥地利", ],
-    "it": ["意大利", "圣马力诺", "梵蒂冈"],
-    "ja": ["日本"],
-    "ru": ["俄罗斯"],
-    "uk": ["乌克兰"],
-    "ar": ["沙特阿拉伯", "埃及", "阿拉伯联合酋长国", "摩洛哥"],
-    "nl":["荷兰"],
-    "no":["挪威"],
-    "sv":["瑞典"],
-    "id":["印度尼西亚"],
-    "vi": ["越南"],
-    "he": ["以色列"],
-    "tr": ["土耳其"],
-    "da": ["丹麦"],
-}
-
-def query_ip_country(client_ip):
-    """
-    Query IP address geo information with robust error handling
-    
-    Features:
-    - 3 second timeout limit
-    - Comprehensive error handling 
-    - Automatic fallback to default values
-    - Cache mechanism to avoid repeated queries
-    
-    Returns:
-        dict: {"country": str, "region": str, "city": str}
-    """
-    # Check cache first - no API call for subsequent visits
-    if client_ip in IP_Country_Cache:
-        print(f"Using cached IP data for {client_ip}")
-        return IP_Country_Cache[client_ip]
-    
-    # Validate IP address
-    if not client_ip or client_ip in ["127.0.0.1", "localhost", "::1"]:
-        print(f"Invalid or local IP address: {client_ip}, using default")
-        default_geo = {"country": "Unknown", "region": "Unknown", "city": "Unknown"}
-        IP_Country_Cache[client_ip] = default_geo
-        return default_geo
-    
-    # First time visit - query API with robust error handling
-    print(f"Querying IP geolocation for {client_ip}...")
-    
-    try:
-        import requests
-        from requests.exceptions import Timeout, ConnectionError, RequestException
-        
-        api_url = f"https://api.vore.top/api/IPdata?ip={client_ip}"
-        
-        # Make request with 3 second timeout
-        response = requests.get(api_url, timeout=3)
-        
-        if response.status_code == 200:
-            data = response.json()
-            if data.get("code") == 200 and "ipdata" in data:
-                ipdata = data["ipdata"]
-                geo_info = {
-                    "country": ipdata.get("info1", "Unknown"),
-                    "region": ipdata.get("info2", "Unknown"),
-                    "city": ipdata.get("info3", "Unknown")
-                }
-                IP_Country_Cache[client_ip] = geo_info
-                print(f"Successfully detected location for {client_ip}: {geo_info['country']}")
-                return geo_info
-            else:
-                print(f"API returned invalid data for {client_ip}: {data}")
-        else:
-            print(f"API request failed with status {response.status_code} for {client_ip}")
-        
-    except Timeout:
-        print(f"Timeout (>3s) querying IP location for {client_ip}, using default")
-    except ConnectionError:
-        print(f"Network connection error for IP {client_ip}, using default")
-    except RequestException as e:
-        print(f"Request error for IP {client_ip}: {e}, using default")
-    except Exception as e:
-        print(f"Unexpected error querying IP {client_ip}: {e}, using default")
-    
-    # All failures lead here - cache default and return
-    default_geo = {"country": "Unknown", "region": "Unknown", "city": "Unknown"}
-    IP_Country_Cache[client_ip] = default_geo
-    print(f"Cached default location for {client_ip}")
-    return default_geo
-
-def query_ip_country_async(client_ip):
-    """
-    Async version that returns immediately with default, then updates cache in background
-    
-    Returns:
-        tuple: (immediate_lang, geo_info_or_none)
-    """
-    # If already cached, return immediately
-    if client_ip in IP_Country_Cache:
-        geo_info = IP_Country_Cache[client_ip]
-        lang = get_lang_from_country(geo_info["country"])
-        return lang, geo_info
-    
-    # Return default immediately, query in background
-    return "en", None
-
-def get_lang_from_country(country):
-    """
-    Map country name to language code with comprehensive validation
-    
-    Features:
-    - Handles invalid/empty input
-    - Case-insensitive matching
-    - Detailed logging
-    - Always returns valid language code
-    
-    Args:
-        country (str): Country name
-        
-    Returns:
-        str: Language code (always valid, defaults to "en")
-    """
-    # Input validation
-    if not country or not isinstance(country, str) or country.strip() == "":
-        print(f"Invalid country provided: '{country}', defaulting to English")
-        return "en"
-    
-    # Normalize country name
-    country = country.strip()
-    if country.lower() == "unknown":
-        print(f"Unknown country, defaulting to English")
-        return "en"
-    
-    try:
-        # Search in country dictionary with case-sensitive match first
-        for lang, countries in country_dict.items():
-            if country in countries:
-                print(f"Matched country '{country}' to language '{lang}'")
-                return lang
-        
-        # If no exact match, try case-insensitive match
-        country_lower = country.lower()
-        for lang, countries in country_dict.items():
-            for country_variant in countries:
-                if country_variant.lower() == country_lower:
-                    print(f"Case-insensitive match: country '{country}' to language '{lang}'")
-                    return lang
-        
-        # No match found
-        print(f"Country '{country}' not found in country_dict, defaulting to English")
-        return "en"
-        
-    except Exception as e:
-        print(f"Error matching country '{country}': {e}, defaulting to English")
-        return "en"
-
-def get_lang_from_ip(client_ip):
-    """
-    Get language based on IP geolocation with comprehensive error handling
-    
-    Features:
-    - Validates input IP address
-    - Handles all possible exceptions
-    - Always returns a valid language code
-    - Defaults to English on any failure
-    - Includes detailed logging
-    
-    Args:
-        client_ip (str): Client IP address
-        
-    Returns:
-        str: Language code (always valid, defaults to "en")
-    """
-    # Input validation
-    if not client_ip or not isinstance(client_ip, str):
-        print(f"Invalid IP address provided: {client_ip}, defaulting to English")
-        return "en"
-    
-    try:
-        # Query geolocation info (has its own error handling and 3s timeout)
-        geo_info = query_ip_country(client_ip)
-        
-        if not geo_info or not isinstance(geo_info, dict):
-            print(f"No geolocation data for {client_ip}, defaulting to English")
-            return "en"
-        
-        # Extract country with fallback
-        country = geo_info.get("country", "Unknown")
-        if not country or country == "Unknown":
-            print(f"Unknown country for IP {client_ip}, defaulting to English") 
-            return "en"
-        
-        # Map country to language
-        detected_lang = get_lang_from_country(country)
-        
-        # Validate language code
-        if not detected_lang or not isinstance(detected_lang, str) or len(detected_lang) != 2:
-            print(f"Invalid language code '{detected_lang}' for {client_ip}, defaulting to English")
-            return "en"
-        
-        print(f"IP {client_ip} -> Country: {country} -> Language: {detected_lang}")
-        return detected_lang
-        
-    except Exception as e:
-        print(f"Unexpected error getting language from IP {client_ip}: {e}, defaulting to English")
-        return "en"  # Always return a valid language code
-
-def is_restricted_country_ip(client_ip):
-    """
-    Check if IP is from a restricted country
-    
-    Returns:
-        bool: True if from restricted country
-    """
-    geo_info = query_ip_country(client_ip)
-    country = geo_info["country"]
-    return country in RESTRICTED_COUNTRIES
-
-def get_ip_max_limit(client_ip):
-    """
-    Get max usage limit for IP based on country
-    
-    Returns:
-        int: Max usage limit
-    """
-    if is_restricted_country_ip(client_ip):
-        return RESTRICTED_COUNTRY_LIMIT
-    else:
-        return RATE_LIMIT_60
-
-def get_ip_generation_count(client_ip):
-    """
-    Get IP generation count
-    """
-    if client_ip not in IP_Generation_Count:
-        IP_Generation_Count[client_ip] = 0
-    return IP_Generation_Count[client_ip]
-
-def increment_ip_generation_count(client_ip):
-    """
-    Increment IP generation count
-    """
-    if client_ip not in IP_Generation_Count:
-        IP_Generation_Count[client_ip] = 0
-    IP_Generation_Count[client_ip] += 1
-    return IP_Generation_Count[client_ip]
-
-def get_ip_phase(client_ip):
-    """
-    Get current phase for IP
-    
-    Returns:
-        str: 'free', 'rate_limit_1', 'rate_limit_2', 'rate_limit_3', 'blocked'
-    """
-    count = get_ip_generation_count(client_ip)
-    max_limit = get_ip_max_limit(client_ip)
-    
-    # For restricted countries, check if they've reached their limit
-    if is_restricted_country_ip(client_ip):
-        if count >= max_limit:
-            return 'blocked'
-        elif count >= max_limit - 2:  # Last 2 attempts
-            return 'rate_limit_3'
-        elif count >= max_limit - 3:  # 3rd attempt from end
-            return 'rate_limit_2'
-        elif count >= max_limit - 4:  # 4th attempt from end
-            return 'rate_limit_1'
-        else:
-            return 'free'
-    
-    # For normal countries, use standard limits
-    if count < FREE_TRY_N:
-        return 'free'
-    elif count < SLOW_TRY_N:
-        return 'rate_limit_1'        # NSFW blur + 5 minutes 2 images
-    elif count < SLOW2_TRY_N:
-        return 'rate_limit_2'        # NSFW blur + 10 minutes 2 images
-    elif count < max_limit:
-        return 'rate_limit_3'        # NSFW blur + 20 minutes 2 images
-    else:
-        return 'blocked'             # Generation blocked
-
-def check_rate_limit_for_phase(client_ip, phase):
-    """
-    Check rate limit for specific phase
-    
-    Returns:
-        tuple: (is_limited, wait_time_minutes, current_count)
-    """
-    if phase not in ['rate_limit_1', 'rate_limit_2', 'rate_limit_3']:
-        return False, 0, 0
-    
-    # Determine time window
-    if phase == 'rate_limit_1':
-        window_minutes = PHASE_1_WINDOW
-    elif phase == 'rate_limit_2':
-        window_minutes = PHASE_2_WINDOW
-    else:  # rate_limit_3
-        window_minutes = PHASE_3_WINDOW
-    
-    current_time = time.time()
-    window_key = f"{client_ip}_{phase}"
-    
-    # Clean expired records
-    if window_key in IP_Rate_Limit_Track:
-        track_data = IP_Rate_Limit_Track[window_key]
-        # Check if within current time window
-        if current_time - track_data['start_time'] > window_minutes * 60:
-            # Time window expired, reset
-            IP_Rate_Limit_Track[window_key] = {
-                'count': 0,
-                'start_time': current_time,
-                'last_generation': current_time
-            }
-    else:
-        # Initialize
-        IP_Rate_Limit_Track[window_key] = {
-            'count': 0,
-            'start_time': current_time,
-            'last_generation': current_time
-        }
-    
-    track_data = IP_Rate_Limit_Track[window_key]
-    
-    # Check if exceeded limit
-    if track_data['count'] >= MAX_IMAGES_PER_WINDOW:
-        # Calculate remaining wait time
-        elapsed = current_time - track_data['start_time']
-        wait_time = (window_minutes * 60) - elapsed
-        wait_minutes = max(0, wait_time / 60)
-        return True, wait_minutes, track_data['count']
-    
-    return False, 0, track_data['count']
-
-def update_country_stats(client_ip):
-    """
-    Update country usage statistics and print periodically
-    """
-    global Total_Request_Count, Country_Usage_Stats
-    
-    # Get country info
-    geo_info = IP_Country_Cache.get(client_ip, {"country": "Unknown", "region": "Unknown", "city": "Unknown"})
-    country = geo_info["country"]
-    
-    # Update country stats
-    if country not in Country_Usage_Stats:
-        Country_Usage_Stats[country] = 0
-    Country_Usage_Stats[country] += 1
-    
-    # Increment total request counter
-    Total_Request_Count += 1
-    
-    # Print stats every N requests
-    if Total_Request_Count % PRINT_STATS_INTERVAL == 0:
-        print("\n" + "="*60)
-        print(f"📊 国家使用统计 (总请求数: {Total_Request_Count})")
-        print("="*60)
-        
-        # Sort by usage count (descending)
-        sorted_stats = sorted(Country_Usage_Stats.items(), key=lambda x: x[1], reverse=True)
-        
-        for country_name, count in sorted_stats:
-            percentage = (count / Total_Request_Count) * 100
-            print(f"  {country_name}: {count} 次 ({percentage:.1f}%)")
-        
-        print("="*60 + "\n")
-
-def record_generation_attempt(client_ip, phase):
-    """
-    Record generation attempt
-    """
-    # Increment total count
-    increment_ip_generation_count(client_ip)
-    
-    # Update country statistics
-    update_country_stats(client_ip)
-    
-    # Record time window count
-    if phase in ['rate_limit_1', 'rate_limit_2', 'rate_limit_3']:
-        window_key = f"{client_ip}_{phase}"
-        current_time = time.time()
-        
-        if window_key in IP_Rate_Limit_Track:
-            IP_Rate_Limit_Track[window_key]['count'] += 1
-            IP_Rate_Limit_Track[window_key]['last_generation'] = current_time
-        else:
-            IP_Rate_Limit_Track[window_key] = {
-                'count': 1,
-                'start_time': current_time,
-                'last_generation': current_time
-            }
-
-def apply_gaussian_blur_to_image_url(image_url, blur_strength=50):
-    """
-    Apply Gaussian blur to image URL
-    
-    Args:
-        image_url (str): Original image URL
-        blur_strength (int): Blur strength, default 50 (heavy blur)
-        
-    Returns:
-        PIL.Image: Blurred PIL Image object
-    """
-    try:
-        import requests
-        from PIL import Image, ImageFilter
-        import io
-        
-        # Download image
-        response = requests.get(image_url, timeout=30)
-        if response.status_code != 200:
-            return None
-        
-        # Convert to PIL Image
-        image_data = io.BytesIO(response.content)
-        image = Image.open(image_data)
-        
-        # Apply heavy Gaussian blur
-        blurred_image = image.filter(ImageFilter.GaussianBlur(radius=blur_strength))
-        
-        return blurred_image
-        
-    except Exception as e:
-        print(f"⚠️ Failed to apply Gaussian blur: {e}")
-        return None
-
-# Initialize NSFW detector (download from Hugging Face)
-try:
-    nsfw_detector = NSFWDetector()  # Auto download falconsai_yolov9_nsfw_model_quantized.pt from Hugging Face
-    print("✅ NSFW detector initialized successfully")
-except Exception as e:
-    print(f"❌ NSFW detector initialization failed: {e}")
-    nsfw_detector = None
-
-def edit_image_interface(input_image, prompt, lang, request: gr.Request, progress=gr.Progress()):
-    """
-    Interface function for processing image editing with phase-based limitations
-    """
-    # 默认禁用“Use as Input”按钮，待成功生成后再开启
-    use_as_input_state = gr.update(interactive=False)
-    try:
-        # Extract user IP
-        client_ip = request.client.host
-        x_forwarded_for = dict(request.headers).get('x-forwarded-for')
-        if x_forwarded_for:
-            client_ip = x_forwarded_for   
-        if client_ip not in IP_Dict:
-            IP_Dict[client_ip] = 0
-        IP_Dict[client_ip] += 1
-
-        if input_image is None:
-            return None, t("error_upload_first", lang), gr.update(visible=False), use_as_input_state
-        
-        if not prompt or prompt.strip() == "":
-            return None, t("error_enter_prompt", lang), gr.update(visible=False), use_as_input_state
-        
-        # Check if prompt length is greater than 3 characters
-        if len(prompt.strip()) <= 3:
-            return None, t("error_prompt_too_short", lang), gr.update(visible=False), use_as_input_state
-    except Exception as e:
-        print(f"⚠️ Unexpected error: {e}", flush=True)
-        return None, t("error_processing_failed", lang), gr.update(visible=False), use_as_input_state
-    
-    # Concurrency guard: block if there is an active task within last 3 minutes
-    try:
-        now_ts = time.time()
-        active_info = Active_Tasks.get(client_ip)
-        if active_info:
-            start_ts = active_info.get("start", 0)
-            if now_ts - start_ts <= 180:
-                return None, "You already have a task in progress. Please wait for it to finish before submitting a new one.", gr.update(visible=False, value=None), use_as_input_state
-            else:
-                # Cleanup stale record
-                Active_Tasks.pop(client_ip, None)
-    except Exception as e:
-        print(f"⚠️ Concurrency guard check failed: {e}")
-    
-    # Get user current phase
-    current_phase = get_ip_phase(client_ip)
-    current_count = get_ip_generation_count(client_ip)
-    geo_info = IP_Country_Cache.get(client_ip, {"country": "Unknown", "region": "Unknown", "city": "Unknown"})
-    is_restricted = is_restricted_country_ip(client_ip)
-    
-    print(f"📊 User phase info - IP: {client_ip}, Location: {geo_info['country']}/{geo_info['region']}/{geo_info['city']}, Phase: {current_phase}, Count: {current_count}, Restricted: {is_restricted}")
-    
-    # Check if user reached the like button tip threshold
-    # For restricted countries, show like tip from the first attempt
-    show_like_tip = (current_count >= 1) if is_restricted else (current_count >= TIP_TRY_N)
-    
-    # Check if completely blocked
-    if current_phase == 'blocked':
-        # Generate blocked limit button with different URL for restricted countries
-        if is_restricted or lang in ["hi", "ru", "zh"]:
-            blocked_url = GOOGLE_GEMINI_URL
-        else:
-            blocked_url = 'https://omnicreator.net/#generator'
-            
-        blocked_button_html = f"""
-        <div style='display: flex; justify-content: center; gap: 15px; margin: 10px 0 5px 0; padding: 0px;'>
-            <a href='{blocked_url}' target='_blank' style='
-                display: inline-flex; 
-                align-items: center;
-                justify-content: center;
-                padding: 16px 32px; 
-                background: linear-gradient(135deg, #e74c3c 0%, #c0392b 100%);
-                color: white; 
-                text-decoration: none; 
-                border-radius: 12px; 
-                font-weight: 600;
-                font-size: 16px;
-                text-align: center;
-                min-width: 200px;
-                box-shadow: 0 4px 15px rgba(231, 76, 60, 0.4);
-                transition: all 0.3s ease;
-                border: none;
-            '>&#128640; Unlimited Generation</a>
-        </div>
-        """
-        
-        # Use same message for all users to avoid discrimination perception
-        blocked_message = t("error_free_limit_reached", lang)
-        
-        return None, blocked_message, gr.update(value=blocked_button_html, visible=True), use_as_input_state
-    
-    # Check rate limit (applies to rate_limit phases)
-    if current_phase in ['rate_limit_1', 'rate_limit_2', 'rate_limit_3']:
-        is_limited, wait_minutes, window_count = check_rate_limit_for_phase(client_ip, current_phase)
-        if is_limited:
-            wait_minutes_int = int(wait_minutes) + 1
-            # Generate rate limit button with different URL for restricted countries
-            if is_restricted or lang in ["hi", "ru", "zh"]:
-                rate_limit_url = GOOGLE_GEMINI_URL
-            else:
-                rate_limit_url = 'https://omnicreator.net/#generator'
-                
-            rate_limit_button_html = f"""
-            <div style='display: flex; justify-content: center; gap: 15px; margin: 10px 0 5px 0; padding: 0px;'>
-                <a href='{rate_limit_url}' target='_blank' style='
-                    display: inline-flex; 
-                    align-items: center;
-                    justify-content: center;
-                    padding: 16px 32px; 
-                    background: linear-gradient(135deg, #f39c12 0%, #e67e22 100%);
-                    color: white; 
-                    text-decoration: none; 
-                    border-radius: 12px; 
-                    font-weight: 600;
-                    font-size: 16px;
-                    text-align: center;
-                    min-width: 200px;
-                    box-shadow: 0 4px 15px rgba(243, 156, 18, 0.4);
-                    transition: all 0.3s ease;
-                    border: none;
-                '>⏰ Skip Wait - Unlimited Generation</a>
-            </div>
-            """
-            return None, t("error_free_limit_wait", lang).format(wait_minutes_int=wait_minutes_int), gr.update(value=rate_limit_button_html, visible=True), use_as_input_state
-    
-    # Handle NSFW detection based on phase
-    is_nsfw_task = False  # Track if this task involves NSFW content
-    
-    # Skip NSFW detection in free phase
-    if current_phase != 'free' and nsfw_detector is not None and input_image is not None:
-        try:
-            nsfw_result = nsfw_detector.predict_pil_label_only(input_image)
-            
-            if nsfw_result.lower() == "nsfw":
-                is_nsfw_task = True
-                use_as_input_state = gr.update(interactive=False)
-                print(f"🔍 Input NSFW detected in {current_phase} phase: ❌❌❌ {nsfw_result} - IP: {client_ip} (will blur result)")
-            else:
-                print(f"🔍 Input NSFW check passed: ✅✅✅ {nsfw_result} - IP: {client_ip}")
-                
-        except Exception as e:
-            print(f"⚠️ Input NSFW detection failed: {e}")
-            # Allow continuation when detection fails
-    
-    result_url = None
-    status_message = ""
-    use_as_input_state = gr.update(interactive=True)
-    
-    def progress_callback(message):
-        try:
-            nonlocal status_message
-            status_message = message
-            # Add error handling to prevent progress update failure
-            if progress is not None:
-                # Enhanced progress display with better formatting
-                if "Queue:" in message or "tasks ahead" in message:
-                    # Queue status - show with different progress value to indicate waiting
-                    progress(0.1, desc=message)
-                elif "Processing" in message or "AI is processing" in message:
-                    # Processing status
-                    progress(0.7, desc=message)
-                elif "Generating" in message or "Almost done" in message:
-                    # Generation status
-                    progress(0.9, desc=message)
-                else:
-                    # Default status
-                    progress(0.5, desc=message)
-        except Exception as e:
-            print(f"⚠️ Progress update failed: {e}")
-    
-    try:
-        # Determine priority before recording generation attempt
-        # First high_priority_n tasks for each IP get priority=1
-        task_priority = 1 if current_count < high_priority_n else 0
-        
-        # Record active task start (for concurrency guard)
-        Active_Tasks[client_ip] = {"start": time.time()}
-        
-        # Record generation attempt (before actual generation to ensure correct count)
-        record_generation_attempt(client_ip, current_phase)
-        updated_count = get_ip_generation_count(client_ip)
-        
-        print(f"✅ Processing started - IP: {client_ip}, phase: {current_phase}, total count: {updated_count}, priority: {task_priority}, prompt: {prompt.strip()}", flush=True)
-        
-        # Call image editing processing function with priority
-        input_image_url, result_url, message, task_uuid = process_image_edit(input_image, prompt.strip(), None, progress_callback, priority=task_priority, client_ip=client_ip)
-        
-        # Check if HF user limit exceeded
-        if message and message.startswith("HF_LIMIT_EXCEEDED:"):
-            error_message = message.replace("HF_LIMIT_EXCEEDED:", "")
-            # Generate HF limit exceeded button (similar to blocked status)
-            hf_limit_url = 'https://omnicreator.net/#generator'
-            
-            hf_limit_button_html = f"""
-            <div style='display: flex; justify-content: center; gap: 15px; margin: 10px 0 5px 0; padding: 0px;'>
-                <a href='{hf_limit_url}' target='_blank' style='
-                    display: inline-flex; 
-                    align-items: center;
-                    justify-content: center;
-                    padding: 16px 32px; 
-                    background: linear-gradient(135deg, #e74c3c 0%, #c0392b 100%);
-                    color: white; 
-                    text-decoration: none; 
-                    border-radius: 12px; 
-                    font-weight: 600;
-                    font-size: 16px;
-                    text-align: center;
-                    min-width: 200px;
-                    box-shadow: 0 4px 15px rgba(231, 76, 60, 0.4);
-                    transition: all 0.3s ease;
-                    border: none;
-                '>&#128640; Unlimited Generation</a>
-            </div>
-            """
-            
-            # Use translated message or default
-            limit_message = error_message if error_message else t("error_free_limit_reached", lang)
-            
-            return None, limit_message, gr.update(value=hf_limit_button_html, visible=True), use_as_input_state
-        
-        if result_url:
-            print(f"✅ Processing completed successfully - IP: {client_ip}, result_url: {result_url}, task_uuid: {task_uuid}", flush=True)
-            
-            # Detect result image NSFW content (only in rate limit phases)
-            if nsfw_detector is not None and current_phase != 'free':
-                try:
-                    if progress is not None:
-                        progress(0.9, desc=t("status_checking_result", lang))
-                    
-                    is_nsfw, nsfw_error = download_and_check_result_nsfw(result_url, nsfw_detector)
-                    
-                    if nsfw_error:
-                        print(f"⚠️ Result image NSFW detection error - IP: {client_ip}, error: {nsfw_error}")
-                    elif is_nsfw:
-                        is_nsfw_task = True  # Mark task as NSFW
-                        print(f"🔍 Result image NSFW detected in {current_phase} phase: ❌❌❌ - IP: {client_ip} (will blur result)")
-                    else:
-                        print(f"🔍 Result image NSFW check passed: ✅✅✅ - IP: {client_ip}")
-                        
-                except Exception as e:
-                    print(f"⚠️ Result image NSFW detection exception - IP: {client_ip}, error: {str(e)}")
-            
-            # Apply blur if this is an NSFW task in rate limit phases
-            should_blur = False
-            
-            if current_phase in ['rate_limit_1', 'rate_limit_2', 'rate_limit_3'] and is_nsfw_task:
-                should_blur = True
-            
-            # Apply blur processing
-            if should_blur:
-                if progress is not None:
-                    progress(0.95, desc=t("status_applying_filter", lang))
-                
-                blurred_image = apply_gaussian_blur_to_image_url(result_url)
-                if blurred_image is not None:
-                    final_result = blurred_image  # Return PIL Image object
-                    final_message = t("warning_content_filter", lang)
-                    print(f"🔒 Applied Gaussian blur for NSFW content - IP: {client_ip}")
-                else:
-                    # Blur failed, return original URL with warning
-                    final_result = result_url
-                    final_message = t("warning_content_review", lang)
-                
-                # Disable use-as-input when NSFW content is detected
-                use_as_input_state = gr.update(interactive=False)
-                
-                # Generate NSFW button for blurred content with different URL for restricted countries
-                if is_restricted or lang in ["hi", "ru", "zh"]:
-                    nsfw_url = GOOGLE_GEMINI_URL
-                else:
-                    nsfw_url = 'https://omnicreator.net/#generator'
-                    
-                banner_html = """
-                <div style='margin: 14px auto 0; max-width: 640px; background: linear-gradient(120deg, #f0f4ff 0%, #e5edff 50%, #f7fbff 100%); border: 1px solid #cbd5ff; border-radius: 14px; padding: 14px 18px; box-shadow: 0 10px 25px rgba(88, 101, 242, 0.18); text-align: center;'>
-                    <div style='font-size: 15px; font-weight: 800; color: #1f2a44; display: flex; align-items: center; justify-content: center; gap: 8px;'>
-                        🚀 Omni Image Editor 2.0 is live!
-                    </div>
-                    <a href='https://huggingface.co/spaces/selfit-camera/Omni-Image-Editor' target='_blank' style='display: inline-flex; align-items: center; justify-content: center; margin-top: 6px; padding: 10px 18px; background: #5865f2; color: white; border-radius: 10px; font-weight: 800; text-decoration: none; box-shadow: 0 6px 18px rgba(88, 101, 242, 0.35);'>
-                        Try the Hugging Face Space demo (free)
-                    </a>
-                    <div style='font-size: 13px; color: #4a5568; margin-top: 6px; font-weight: 600;'>This is a free HF Space demo for Omni Image Editor 2.0.</div>
-                </div>
-                """
-                
-                nsfw_action_buttons_html = f"""
-                <div style='text-align: center; margin: 18px 0 10px 0;'>
-                    <a href='{nsfw_url}' target='_blank' style='
-                        display: inline-flex; 
-                        align-items: center;
-                        justify-content: center;
-                        padding: 16px 32px; 
-                        background: linear-gradient(135deg, #ff6b6b 0%, #feca57 100%);
-                        color: white; 
-                        text-decoration: none; 
-                        border-radius: 12px; 
-                        font-weight: 700;
-                        font-size: 16px;
-                        min-width: 220px;
-                        box-shadow: 0 8px 25px rgba(255, 107, 107, 0.35);
-                        transition: all 0.3s ease;
-                        border: none;
-                    '>🔥 Unlimited Creative Generation</a>
-                </div>
-                {banner_html}
-                """
-                return final_result, final_message, gr.update(value=nsfw_action_buttons_html, visible=True), use_as_input_state
-            else:
-                final_result = result_url
-                final_message = t("status_completed_message", lang).format(message=message)
-            
-            try:
-                if progress is not None:
-                    progress(1.0, desc=t("status_processing_completed", lang))
-            except Exception as e:
-                print(f"⚠️ Final progress update failed: {e}")
-            
-            # Generate action buttons HTML
-            banner_html = """
-            <div style='margin: 14px auto 0; max-width: 640px; background: linear-gradient(120deg, #f0f4ff 0%, #e5edff 50%, #f7fbff 100%); border: 1px solid #cbd5ff; border-radius: 14px; padding: 14px 18px; box-shadow: 0 10px 25px rgba(88, 101, 242, 0.18); text-align: center;'>
-                <div style='font-size: 15px; font-weight: 800; color: #1f2a44; display: flex; align-items: center; justify-content: center; gap: 8px;'>
-                    🚀 Omni Image Editor 2.0 is live!
-                </div>
-                <a href='https://huggingface.co/spaces/selfit-camera/Omni-Image-Editor' target='_blank' style='display: inline-flex; align-items: center; justify-content: center; margin-top: 6px; padding: 10px 18px; background: #5865f2; color: white; border-radius: 10px; font-weight: 800; text-decoration: none; box-shadow: 0 6px 18px rgba(88, 101, 242, 0.35);'>
-                    Try the Hugging Face Space demo (free)
-                </a>
-                <div style='font-size: 13px; color: #4a5568; margin-top: 6px; font-weight: 600;'>This is a free HF Space demo for Omni Image Editor 2.0.</div>
-            </div>
-            """
-            
-            action_buttons_html = ""
-            
-            # 根据 TIP_TRY_N（受限地区从第一次起就触发）展示点赞提示
-            if show_like_tip:
-                action_buttons_html = """
-                <div style='display: flex; justify-content: center; margin: 15px 0 5px 0; padding: 0px;'>
-                    <div style='
-                        display: inline-flex; 
-                        align-items: center;
-                        justify-content: center;
-                        padding: 12px 24px; 
-                        background: linear-gradient(135deg, #7c3aed 0%, #6366f1 100%);
-                        color: white; 
-                        text-decoration: none; 
-                        border-radius: 10px; 
-                        font-weight: 600;
-                        font-size: 14px;
-                        text-align: center;
-                        max-width: 400px;
-                        box-shadow: 0 3px 12px rgba(255, 107, 107, 0.3);
-                        border: none;
-                    '>👉 Click the ❤️ Like button to unlock more free trial attempts!</div>
-                </div>
-                """
-            
-            # Always show the Omni Image Editor 2.0 banner under the result image
-            action_buttons_html = f"{action_buttons_html}{banner_html}"
-            
-            return final_result, final_message, gr.update(value=action_buttons_html, visible=True), use_as_input_state
-        else:
-            print(f"❌ Processing failed - IP: {client_ip}, error: {message}", flush=True)
-            return None, t("error_processing_failed", lang).format(message=message), gr.update(visible=False), use_as_input_state
-            
-    except Exception as e:
-        print(f"❌ Processing exception - IP: {client_ip}, error: {str(e)}")
-        return None, t("error_processing_exception", lang).format(error=str(e)), gr.update(visible=False), use_as_input_state
-    finally:
-        # Task finished (success or failure) — clear active marker to allow next submission immediately
-        Active_Tasks.pop(client_ip, None)
-
-# Create Gradio interface
-def create_app():
-    with gr.Blocks(
-        title="Image Editor 1.0",
-        theme=gr.themes.Soft(),
-        css="""
-        .main-container {
-            max-width: 1200px;
-            margin: 0 auto;
-        }
-        .news-banner-row {
-            margin: 10px auto 15px auto;
-            padding: 0 10px;
-            max-width: 1200px;
-            width: 100% !important;
-        }
-        .news-banner-row .gr-row {
-            display: flex !important;
-            align-items: center !important;
-            width: 100% !important;
-        }
-        .news-banner-row .gr-column:first-child {
-            flex: 1 !important; /* 占据所有剩余空间 */
-            display: flex !important;
-            justify-content: center !important; /* 在其空间内居中 */
-        }
-        .banner-lang-selector {
-            margin-left: auto !important;
-            display: flex !important;
-            justify-content: flex-end !important;
-            align-items: center !important;
-            position: relative !important;
-            z-index: 10 !important;
-        }
-        .banner-lang-selector .gr-dropdown {
-            background: white !important;
-            border: 1px solid #ddd !important;
-            border-radius: 8px !important;
-            padding: 8px 16px !important;
-            font-size: 14px !important;
-            font-weight: 500 !important;
-            color: #333 !important;
-            cursor: pointer !important;
-            box-shadow: 0 2px 4px rgba(0, 0, 0, 0.1) !important;
-            min-width: 140px !important;
-            max-width: 160px !important;
-            transition: all 0.2s ease !important;
-        }
-        .banner-lang-selector .gr-dropdown:hover {
-            border-color: #999 !important;
-            box-shadow: 0 2px 8px rgba(0, 0, 0, 0.15) !important;
-        }
-        @media (max-width: 768px) {
-            .news-banner-row {
-                padding: 0 15px !important;
-            }
-            .news-banner-row .gr-row {
-                display: flex !important;
-                flex-direction: column !important;
-                gap: 10px !important;
-                position: static !important;
-            }
-            .news-banner-row .gr-column:first-child {
-                position: static !important;
-                pointer-events: auto !important;
-            }
-            .banner-lang-selector {
-                margin-left: 0 !important;
-                justify-content: center !important;
-            }
-        }
-        .upload-area {
-            border: 2px dashed #ccc;
-            border-radius: 10px;
-            padding: 20px;
-            text-align: center;
-        }
-        .result-area {
-            margin-top: 20px;
-            padding: 20px;
-            border-radius: 10px;
-            background-color: #f8f9fa;
-        }
-        .use-as-input-btn {
-            margin-top: 10px;
-            width: 100%;
-        }
-        """,
-        # Improve concurrency performance configuration
-        head="""
-        <script>
-        // Reduce client-side state update frequency, avoid excessive SSE connections
-        if (window.gradio) {
-            window.gradio.update_frequency = 2000; // Update every 2 seconds
-        }
-        </script>
-        """
-    ) as app:
-        
-        lang_state = gr.State("en")
-
-        # Main title - centered
-        header_title = gr.HTML(f"""
-        <div style="text-align: center; margin: 20px auto 10px auto; max-width: 800px;">
-            <h1 style="color: #2c3e50; margin: 0; font-size: 3.5em; font-weight: 800; letter-spacing: 3px; text-shadow: 2px 2px 4px rgba(0,0,0,0.1);">
-                {t('header_title', 'en')}
-            </h1>
-        </div>
-        """)
-        
-        with gr.Row(elem_classes=["news-banner-row"]):
-            with gr.Column(scale=1, min_width=400):
-                # Banner is initially visible (will be hidden for zh/hi/ru languages on load)
-                news_banner = gr.HTML(f"""
-                <style>
-                @keyframes breathe {{
-                    0%, 100% {{ transform: scale(1); }}
-                    50% {{ transform: scale(1.02); }}
-                }}
-                .breathing-banner {{
-                    animation: breathe 3s ease-in-out infinite;
-                }}
-                </style>
-                <div class="breathing-banner" style="
-                    background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
-                    margin: 0 auto;
-                    padding: 8px 40px;
-                    border-radius: 20px;
-                    max-width: 800px;
-                    box-shadow: 0 2px 8px rgba(102, 126, 234, 0.3);
-                    text-align: center;
-                    width: fit-content;
-                ">
-                    <span style="color: white; font-weight: 600; font-size: 1.0em;">
-                        🎉 NEW: 
-                        <a href="https://huggingface.co/spaces/selfit-camera/Omni-Image-Editor" target="_blank" style="
-                            color: white; 
-                            text-decoration: none; 
-                            border-bottom: 1px solid rgba(255,255,255,0.5);
-                            transition: all 0.3s ease;
-                            margin: 0 8px;
-                        " onmouseover="this.style.borderBottom='1px solid white'" 
-                           onmouseout="this.style.borderBottom='1px solid rgba(255,255,255,0.5)'">
-                           Image Editor 2.0
-                        </a>
-                        is Online Now ! More free trials, better quality!
-                    </span>
-                </div>
-                """, visible=True)
-            
-            with gr.Column(scale=0, min_width=160, elem_classes=["banner-lang-selector"]):
-                # Lock UI to English only; allow_custom_value avoids Gradio errors if any non-en value is set programmatically
-                lang_dropdown = gr.Dropdown(
-                    choices=[
-                        ("English", "en"),
-                    ],
-                    value="en",
-                    label="🌐",
-                    show_label=True,
-                    interactive=True,
-                    container=False,
-                    allow_custom_value=True,
-                )
-        
-        with gr.Tabs() as tabs:
-            with gr.Tab(t("global_editor_tab", "en")) as global_tab:
-                with gr.Row():
-                    with gr.Column(scale=1):
-                        upload_image_header = gr.Markdown(t("upload_image_header", "en"))
-                        input_image = gr.Image(
-                            label=t("upload_image_label", "en"),
-                            type="pil",
-                            height=512,
-                            elem_classes=["upload-area"]
-                        )
-
-                        editing_instructions_header = gr.Markdown(t("editing_instructions_header", "en"))
-                        prompt_input = gr.Textbox(
-                            label=t("prompt_input_label", "en"),
-                            placeholder=t("prompt_input_placeholder", "en"),
-                            lines=3,
-                            max_lines=5
-                        )
-                        
-                        edit_button = gr.Button(
-                            t("start_editing_button", "en"),
-                            variant="primary",
-                            size="lg"
-                        )
-                    
-                    with gr.Column(scale=1):
-                        editing_result_header = gr.Markdown(t("editing_result_header", "en"))
-                        output_image = gr.Image(
-                            label=t("output_image_label", "en"),
-                            height=320,
-                            elem_classes=["result-area"]
-                        )
-                        
-                        use_as_input_btn = gr.Button(
-                            t("use_as_input_button", "en"),
-                            variant="secondary",
-                            size="sm",
-                            elem_classes=["use-as-input-btn"]
-                        )
-                        
-                        status_output = gr.Textbox(
-                            label=t("status_output_label", "en"),
-                            lines=2,
-                            max_lines=3,
-                            interactive=False
-                        )
-                        
-                        action_buttons = gr.HTML(visible=False)
-                
-                prompt_examples_header = gr.Markdown(t("prompt_examples_header", "en"))
-                with gr.Row():
-                    example_prompts = [
-                        "Set the background to a grand opera stage with red curtains",
-                        "Change the outfit into a traditional Chinese hanfu with flowing sleeves",
-                        "Give the character blue dragon-like eyes with glowing pupils",
-                        "Change lighting to soft dreamy pastel glow",
-                        "Change pose to sitting cross-legged on the ground"
-                    ]
-                    
-                    for prompt in example_prompts:
-                        gr.Button(
-                            prompt,
-                            size="sm"
-                        ).click(
-                            lambda p=prompt: p,
-                            outputs=prompt_input
-                        )
-                
-                edit_button.click(
-                    fn=edit_image_interface,
-                    inputs=[input_image, prompt_input, lang_state],
-                    outputs=[output_image, status_output, action_buttons, use_as_input_btn],
-                    show_progress=True,
-                    concurrency_limit=20
-                )
-                
-                def simple_use_as_input(output_img):
-                    if output_img is not None:
-                        return output_img
-                    return None
-                
-                use_as_input_btn.click(
-                    fn=simple_use_as_input,
-                    inputs=[output_image],
-                    outputs=[input_image]
-                )
-        
-        # SEO Content Section
-        seo_html = gr.HTML()
-
-        def get_seo_html(lang):
-            # 中文、印度语、俄语不显示SEO部分
-            if lang in ["zh", "hi", "ru"]:
-                return ""
-            
-            return f"""
-    <div style="width: 100%; margin: 50px 0; padding: 0 20px;">
-        
-        <div style="text-align: center; background: linear-gradient(135deg, #667eea 0%, #764ba2 100%); color: white; padding: 40px; border-radius: 20px; margin: 40px 0;">
-            <h2 style="margin: 0 0 20px 0; font-size: 2.2em; font-weight: 700;">
-                &#127912; {t('seo_unlimited_title', lang)}
-            </h2>
-            <p style="margin: 0 0 25px 0; font-size: 1.2em; opacity: 0.95; line-height: 1.6;">
-                {t('seo_unlimited_desc', lang)}
-            </p>
-            
-            <div style="display: flex; justify-content: center; gap: 25px; flex-wrap: wrap; margin: 30px 0;">
-                <a href="https://omnicreator.net/#generator" target="_blank" style="
-                    display: inline-flex; 
-                    align-items: center;
-                    justify-content: center;
-                    padding: 20px 40px; 
-                    background: linear-gradient(135deg, #ff6b6b 0%, #feca57 100%);
-                    color: white; 
-                    text-decoration: none; 
-                    border-radius: 15px; 
-                    font-weight: 700;
-                    font-size: 18px;
-                    text-align: center;
-                    min-width: 250px;
-                    box-shadow: 0 8px 25px rgba(255, 107, 107, 0.4);
-                    transition: all 0.3s ease;
-                    border: none;
-                    transform: scale(1);
-                " onmouseover="this.style.transform='scale(1.05)'" onmouseout="this.style.transform='scale(1)'">
-                    &#128640; {t('seo_unlimited_button', lang)}
-                </a>
-                
-            </div>
-            
-            <p style="color: rgba(255,255,255,0.9); font-size: 1em; margin: 20px 0 0 0;">
-                {t('seo_unlimited_footer', lang)}
-            </p>
-        </div>
-
-        <div style="text-align: center; margin: 25px auto; background: linear-gradient(135deg, #f5f7fa 0%, #c3cfe2 100%); padding: 35px; border-radius: 20px; box-shadow: 0 10px 30px rgba(0,0,0,0.1);">
-            <h2 style="color: #2c3e50; margin: 0 0 20px 0; font-size: 1.9em; font-weight: 700;">
-                &#11088; {t('seo_professional_title', lang)}
-            </h2>
-            <p style="color: #555; font-size: 1.1em; line-height: 1.6; margin: 0 0 20px 0; padding: 0 20px;">
-                {t('seo_professional_desc', lang)}
-            </p>
-        </div>
-
-        <div style="display: grid; grid-template-columns: repeat(auto-fit, minmax(300px, 1fr)); gap: 25px; margin: 40px 0;">
-            
-            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #e74c3c;">
-                <h3 style="color: #e74c3c; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
-                    &#127919; {t('seo_feature1_title', lang)}
-                </h3>
-                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
-                    {t('seo_feature1_desc', lang)}
-                </p>
-            </div>
-
-            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #3498db;">
-                <h3 style="color: #3498db; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
-                    🔓 {t('seo_feature2_title', lang)}
-                </h3>
-                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
-                    {t('seo_feature2_desc', lang)}
-                </p>
-            </div>
-
-            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #27ae60;">
-                <h3 style="color: #27ae60; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
-                    &#9889; {t('seo_feature3_title', lang)}
-                </h3>
-                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
-                    {t('seo_feature3_desc', lang)}
-                </p>
-            </div>
-
-            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #9b59b6;">
-                <h3 style="color: #9b59b6; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
-                    &#127912; {t('seo_feature4_title', lang)}
-                </h3>
-                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
-                    {t('seo_feature4_desc', lang)}
-                </p>
-            </div>
-
-            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #f39c12;">
-                <h3 style="color: #f39c12; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
-                    &#128142; {t('seo_feature5_title', lang)}
-                </h3>
-                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
-                    {t('seo_feature5_desc', lang)}
-                </p>
-            </div>
-
-            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #34495e;">
-                <h3 style="color: #34495e; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
-                    🌍 {t('seo_feature6_title', lang)}
-                </h3>
-                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
-                    {t('seo_feature6_desc', lang)}
-                </p>
-            </div>
-
-        </div>
-
-        <div style="background: linear-gradient(135deg, #ff9a9e 0%, #fecfef 50%, #fecfef 100%); padding: 30px; border-radius: 15px; margin: 40px 0;">
-            <h3 style="color: #8b5cf6; text-align: center; margin: 0 0 25px 0; font-size: 1.5em; font-weight: 700;">
-                &#128161; {t('seo_protips_title', lang)}
-            </h3>
-            <div style="display: grid; grid-template-columns: repeat(auto-fit, minmax(300px, 1fr)); gap: 18px;">
-                
-                <div style="background: rgba(255,255,255,0.85); padding: 18px; border-radius: 12px;">
-                    <strong style="color: #8b5cf6; font-size: 1.1em;">📝 {t('seo_protip1_title', lang)}</strong>
-                    <p style="color: #555; margin: 5px 0 0 0; line-height: 1.5;">{t('seo_protip1_desc', lang)}</p>
-                </div>
-
-                <div style="background: rgba(255,255,255,0.85); padding: 18px; border-radius: 12px;">
-                    <strong style="color: #8b5cf6; font-size: 1.1em;">&#127919; {t('seo_protip2_title', lang)}</strong>
-                    <p style="color: #555; margin: 5px 0 0 0; line-height: 1.5;">{t('seo_protip2_desc', lang)}</p>
-                </div>
-
-                <div style="background: rgba(255,255,255,0.85); padding: 18px; border-radius: 12px;">
-                    <strong style="color: #8b5cf6; font-size: 1.1em;">&#9889; {t('seo_protip3_title', lang)}</strong>
-                    <p style="color: #555; margin: 5px 0 0 0; line-height: 1.5;">{t('seo_protip3_desc', lang)}</p>
-                </div>
-
-                <div style="background: rgba(255,255,255,0.85); padding: 18px; border-radius: 12px;">
-                    <strong style="color: #8b5cf6; font-size: 1.1em;">&#128444; {t('seo_protip4_title', lang)}</strong>
-                    <p style="color: #555; margin: 5px 0 0 0; line-height: 1.5;">{t('seo_protip4_desc', lang)}</p>
-                </div>
-
-            </div>
-        </div>
-
-        <div style="text-align: center; margin: 25px auto; background: linear-gradient(135deg, #a8edea 0%, #fed6e3 100%); padding: 35px; border-radius: 20px; box-shadow: 0 10px 30px rgba(0,0,0,0.1);">
-            <h2 style="color: #2c3e50; margin: 0 0 20px 0; font-size: 1.8em; font-weight: 700;">
-                &#128640; {t('seo_needs_title', lang)}
-            </h2>
-            <div style="display: grid; grid-template-columns: repeat(auto-fit, minmax(200px, 1fr)); gap: 20px; margin: 25px 0; text-align: left;">
-                
-                <div style="background: rgba(255,255,255,0.8); padding: 20px; border-radius: 12px;">
-                    <h4 style="color: #e74c3c; margin: 0 0 10px 0;">🎨 {t('seo_needs_art_title', lang)}</h4>
-                    <ul style="color: #555; margin: 0; padding-left: 18px; line-height: 1.6;">
-                        <li>{t('seo_needs_art_item1', lang)}</li>
-                        <li>{t('seo_needs_art_item2', lang)}</li>
-                        <li>{t('seo_needs_art_item3', lang)}</li>
-                        <li>{t('seo_needs_art_item4', lang)}</li>
-                    </ul>
-                </div>
-
-                <div style="background: rgba(255,255,255,0.8); padding: 20px; border-radius: 12px;">
-                    <h4 style="color: #3498db; margin: 0 0 10px 0;">📸 {t('seo_needs_photo_title', lang)}</h4>
-                    <ul style="color: #555; margin: 0; padding-left: 18px; line-height: 1.6;">
-                        <li>{t('seo_needs_photo_item1', lang)}</li>
-                        <li>{t('seo_needs_photo_item2', lang)}</li>
-                        <li>{t('seo_needs_photo_item3', lang)}</li>
-                        <li>{t('seo_needs_photo_item4', lang)}</li>
-                    </ul>
-                </div>
-
-                <div style="background: rgba(255,255,255,0.8); padding: 20px; border-radius: 12px;">
-                    <h4 style="color: #27ae60; margin: 0 0 10px 0;">🛍️ {t('seo_needs_ecom_title', lang)}</h4>
-                    <ul style="color: #555; margin: 0; padding-left: 18px; line-height: 1.6;">
-                        <li>{t('seo_needs_ecom_item1', lang)}</li>
-                        <li>{t('seo_needs_ecom_item2', lang)}</li>
-                        <li>{t('seo_needs_ecom_item3', lang)}</li>
-                        <li>{t('seo_needs_ecom_item4', lang)}</li>
-                    </ul>
-                </div>
-
-                <div style="background: rgba(255,255,255,0.8); padding: 20px; border-radius: 12px;">
-                    <h4 style="color: #9b59b6; margin: 0 0 10px 0;">📱 {t('seo_needs_social_title', lang)}</h4>
-                    <ul style="color: #555; margin: 0; padding-left: 18px; line-height: 1.6;">
-                        <li>{t('seo_needs_social_item1', lang)}</li>
-                        <li>{t('seo_needs_social_item2', lang)}</li>
-                        <li>{t('seo_needs_social_item3', lang)}</li>
-                        <li>{t('seo_needs_social_item4', lang)}</li>
-                    </ul>
-                </div>
-
-            </div>
-        </div>
-
-    </div>
-    """
-        
-        all_ui_components = [
-            header_title, news_banner, 
-            global_tab, upload_image_header, input_image, editing_instructions_header, prompt_input, edit_button,
-            editing_result_header, output_image, use_as_input_btn, status_output, prompt_examples_header,
-            seo_html,
-        ]
-
-        def update_ui_lang(lang):
-            # Hide banner for zh, hi, ru languages
-            show_banner = lang not in ["zh", "hi", "ru"]
-            
-            return {
-                header_title: gr.update(value=f"""
-                    <div style="text-align: center; margin: 20px auto 10px auto; max-width: 800px;">
-                        <h1 style="color: #2c3e50; margin: 0; font-size: 3.5em; font-weight: 800; letter-spacing: 3px; text-shadow: 2px 2px 4px rgba(0,0,0,0.1);">
-                            {t('header_title', lang)}
-                        </h1>
-                    </div>"""),
-                news_banner: gr.update(visible=show_banner),
-                global_tab: gr.update(label=t("global_editor_tab", lang)),
-                upload_image_header: gr.update(value=t("upload_image_header", lang)),
-                input_image: gr.update(label=t("upload_image_label", lang)),
-                editing_instructions_header: gr.update(value=t("editing_instructions_header", lang)),
-                prompt_input: gr.update(label=t("prompt_input_label", lang), placeholder=t("prompt_input_placeholder", lang)),
-                edit_button: gr.update(value=t("start_editing_button", lang)),
-                editing_result_header: gr.update(value=t("editing_result_header", lang)),
-                output_image: gr.update(label=t("output_image_label", lang)),
-                use_as_input_btn: gr.update(value=t("use_as_input_button", lang)),
-                status_output: gr.update(label=t("status_output_label", lang)),
-                prompt_examples_header: gr.update(value=t("prompt_examples_header", lang)),
-                seo_html: gr.update(value=get_seo_html(lang)),
-            }
-
-        def on_lang_change(lang):
-            # Force UI to stay in English regardless of dropdown value
-            return "en", *update_ui_lang("en").values()
-            
-        lang_dropdown.change(
-            on_lang_change,
-            inputs=[lang_dropdown],
-            outputs=[lang_state] + all_ui_components
-        )
-
-        # IP query state for async loading
-        ip_query_state = gr.State({"status": "pending", "ip": None, "lang": "en"})
-        
-        def on_load_immediate(request: gr.Request):
-            """
-            Load page with language based on robust IP detection
-            
-            Features:
-            - Multiple fallback layers for IP extraction
-            - Comprehensive error handling
-            - Always returns valid language (defaults to English)
-            - Detailed logging for debugging
-            """
-            # Extract client IP with multiple fallback methods
-            client_ip = None
-            try:
-                # Primary method: direct client host
-                client_ip = request.client.host
-                
-                # Secondary method: check forwarded headers  
-                headers = dict(request.headers) if hasattr(request, 'headers') else {}
-                x_forwarded_for = headers.get('x-forwarded-for') or headers.get('X-Forwarded-For')
-                if x_forwarded_for:
-                    # Take first IP from comma-separated list
-                    client_ip = x_forwarded_for.split(',')[0].strip()
-                
-                # Alternative headers
-                if not client_ip or client_ip in ["127.0.0.1", "localhost"]:
-                    client_ip = headers.get('x-real-ip') or headers.get('X-Real-IP') or client_ip
-                
-            except Exception as e:
-                print(f"Error extracting client IP: {e}, using default")
-                client_ip = "unknown"
-            
-            # Validate extracted IP
-            if not client_ip:
-                client_ip = "unknown"
-            
-            print(f"Loading page for IP: {client_ip}")
-            
-            # Determine language with robust error handling
-            try:
-                # Check if IP is already cached (second+ visit)
-                if client_ip in IP_Country_Cache:
-                    # Use cached data - but force English UI
-                    cached_lang = "en"
-                    print(f"Using cached language (forced to en) for IP: {client_ip}")
-                    query_state = {"ip": client_ip, "cached": True}
-                    return cached_lang, cached_lang, query_state, *update_ui_lang(cached_lang).values()
-                
-                # First visit: Query IP and determine language (max 3s timeout built-in)
-                print(f"First visit - detecting language for IP: {client_ip}")
-                # Always force English UI even if detection yields another language
-                detected_lang = "en"
-                
-                print(f"First visit - Final language forced to: {detected_lang} for IP: {client_ip}")
-                query_state = {"ip": client_ip, "cached": False}
-                return detected_lang, detected_lang, query_state, *update_ui_lang(detected_lang).values()
-                
-            except Exception as e:
-                # Ultimate fallback - always works
-                print(f"Critical error in language detection for {client_ip}: {e}")
-                print("Using English as ultimate fallback")
-                query_state = {"ip": client_ip or "unknown", "cached": False, "error": str(e)}
-                return "en", "en", query_state, *update_ui_lang("en").values()
-
-
-        app.load(
-            on_load_immediate,
-            inputs=None,
-            outputs=[lang_state, lang_dropdown, ip_query_state] + all_ui_components,
-        )
-
-    return app
-
-if __name__ == "__main__":
-    app = create_app()
-    # Improve queue configuration to handle high concurrency and prevent SSE connection issues
-    app.queue(
-        default_concurrency_limit=20,  # Default concurrency limit
-        max_size=50,  # Maximum queue size
-        api_open=False  # Close API access to reduce resource consumption
-    )
-    app.launch(
-        server_name="0.0.0.0",
-        show_error=True,  # Show detailed error information
-        quiet=False,  # Keep log output
-        max_threads=40,  # Increase thread pool size
-        height=800,
-        favicon_path=None  # Reduce resource loading
-    )

__lib__/i18n/__init__.py

@@ -1,36 +0,0 @@
-"""
-i18n loader for encrypted translation files
-"""
-import sys
-import importlib.util
-from pathlib import Path
-
-def load_pyc_module(module_name, pyc_path):
-    """Load a .pyc module using importlib"""
-    spec = importlib.util.spec_from_file_location(module_name, pyc_path)
-    if spec is None or spec.loader is None:
-        raise ImportError(f"Cannot load module {module_name} from {pyc_path}")
-    module = importlib.util.module_from_spec(spec)
-    sys.modules[module_name] = module
-    spec.loader.exec_module(module)
-    return module
-
-def load_translations():
-    """Load all encrypted translation files"""
-    translations = {}
-    i18n_dir = Path(__file__).parent
-    
-    # List all .pyc files in i18n directory
-    for pyc_file in i18n_dir.glob("*.pyc"):
-        lang = pyc_file.stem  # Get language code from filename
-        try:
-            module = load_pyc_module(f"i18n_{lang}", pyc_file)
-            if hasattr(module, 'data'):
-                translations[lang] = module.data
-        except Exception as e:
-            print(f"Failed to load {pyc_file.name}: {e}")
-    
-    return translations
-
-# Auto-load translations when module is imported
-translations = load_translations()

app.py

@@ -1,60 +1,1448 @@
-"""
-Minimal app loader for ImageEditSpace
-This app loads the compiled, obfuscated modules from __lib__
-"""
-import sys
-from pathlib import Path
-import importlib.util
-
-# Add __lib__ to path to import compiled modules
-lib_dir = Path(__file__).parent / "__lib__"
-if not lib_dir.exists():
-    raise RuntimeError(f"Compiled library directory not found: {lib_dir}")
-
-sys.path.insert(0, str(lib_dir))
-
-def load_pyc_module(module_name, pyc_path):
-    """Load a .pyc module using importlib"""
-    spec = importlib.util.spec_from_file_location(module_name, pyc_path)
-    if spec is None or spec.loader is None:
-        raise ImportError(f"Cannot load module {module_name} from {pyc_path}")
-    module = importlib.util.module_from_spec(spec)
-    sys.modules[module_name] = module
-    spec.loader.exec_module(module)
-    return module
+import gradio as gr
+import threading
+import os
+import shutil
+import tempfile
+import time
+from util import process_image_edit, process_local_image_edit, download_and_check_result_nsfw
+from nfsw import NSFWDetector
 
+# Configuration parameters
+
+TIP_TRY_N = 8       # Show like button tip after 12 tries
+FREE_TRY_N = 20        # Free phase: first 15 tries without restrictions
+SLOW_TRY_N = 25       # Slow phase start: 25 tries
+SLOW2_TRY_N = 32        # Slow phase start: 32 tries
+RATE_LIMIT_60 = 40     # Full restriction: blocked after 40 tries
+
+# Time window configuration (minutes)
+PHASE_1_WINDOW = 5     # 15-25 tries: 5 minutes
+PHASE_2_WINDOW = 10    # 25-32 tries: 10 minutes
+PHASE_3_WINDOW = 20    # 32-40 tries: 20 minutes
+MAX_IMAGES_PER_WINDOW = 2  # Max images per time window
+
+IP_Dict = {}
+# IP generation statistics and time window tracking
+IP_Generation_Count = {}  # Record total generation count for each IP
+IP_Rate_Limit_Track = {}  # Record generation count and timestamp in current time window for each IP
+
+def get_ip_generation_count(client_ip):
+    """
+    Get IP generation count
+    """
+    if client_ip not in IP_Generation_Count:
+        IP_Generation_Count[client_ip] = 0
+    return IP_Generation_Count[client_ip]
+
+def increment_ip_generation_count(client_ip):
+    """
+    Increment IP generation count
+    """
+    if client_ip not in IP_Generation_Count:
+        IP_Generation_Count[client_ip] = 0
+    IP_Generation_Count[client_ip] += 1
+    return IP_Generation_Count[client_ip]
+
+def get_ip_phase(client_ip):
+    """
+    Get current phase for IP
+    
+    Returns:
+        str: 'free', 'rate_limit_1', 'rate_limit_2', 'rate_limit_3', 'blocked'
+    """
+    count = get_ip_generation_count(client_ip)
+    
+    if count < FREE_TRY_N:
+        return 'free'
+    elif count < SLOW_TRY_N:
+        return 'rate_limit_1'        # NSFW blur + 5 minutes 2 images
+    elif count < SLOW2_TRY_N:
+        return 'rate_limit_2'        # NSFW blur + 10 minutes 2 images
+    elif count < RATE_LIMIT_60:
+        return 'rate_limit_3'        # NSFW blur + 20 minutes 2 images
+    else:
+        return 'blocked'             # Generation blocked
+
+def check_rate_limit_for_phase(client_ip, phase):
+    """
+    Check rate limit for specific phase
+    
+    Returns:
+        tuple: (is_limited, wait_time_minutes, current_count)
+    """
+    if phase not in ['rate_limit_1', 'rate_limit_2', 'rate_limit_3']:
+        return False, 0, 0
+    
+    # Determine time window
+    if phase == 'rate_limit_1':
+        window_minutes = PHASE_1_WINDOW
+    elif phase == 'rate_limit_2':
+        window_minutes = PHASE_2_WINDOW
+    else:  # rate_limit_3
+        window_minutes = PHASE_3_WINDOW
+    
+    current_time = time.time()
+    window_key = f"{client_ip}_{phase}"
+    
+    # Clean expired records
+    if window_key in IP_Rate_Limit_Track:
+        track_data = IP_Rate_Limit_Track[window_key]
+        # Check if within current time window
+        if current_time - track_data['start_time'] > window_minutes * 60:
+            # Time window expired, reset
+            IP_Rate_Limit_Track[window_key] = {
+                'count': 0,
+                'start_time': current_time,
+                'last_generation': current_time
+            }
+    else:
+        # Initialize
+        IP_Rate_Limit_Track[window_key] = {
+            'count': 0,
+            'start_time': current_time,
+            'last_generation': current_time
+        }
+    
+    track_data = IP_Rate_Limit_Track[window_key]
+    
+    # Check if exceeded limit
+    if track_data['count'] >= MAX_IMAGES_PER_WINDOW:
+        # Calculate remaining wait time
+        elapsed = current_time - track_data['start_time']
+        wait_time = (window_minutes * 60) - elapsed
+        wait_minutes = max(0, wait_time / 60)
+        return True, wait_minutes, track_data['count']
+    
+    return False, 0, track_data['count']
+
+def record_generation_attempt(client_ip, phase):
+    """
+    Record generation attempt
+    """
+    # Increment total count
+    increment_ip_generation_count(client_ip)
+    
+    # Record time window count
+    if phase in ['rate_limit_1', 'rate_limit_2', 'rate_limit_3']:
+        window_key = f"{client_ip}_{phase}"
+        current_time = time.time()
+        
+        if window_key in IP_Rate_Limit_Track:
+            IP_Rate_Limit_Track[window_key]['count'] += 1
+            IP_Rate_Limit_Track[window_key]['last_generation'] = current_time
+        else:
+            IP_Rate_Limit_Track[window_key] = {
+                'count': 1,
+                'start_time': current_time,
+                'last_generation': current_time
+            }
+
+def apply_gaussian_blur_to_image_url(image_url, blur_strength=50):
+    """
+    Apply Gaussian blur to image URL
+    
+    Args:
+        image_url (str): Original image URL
+        blur_strength (int): Blur strength, default 50 (heavy blur)
+        
+    Returns:
+        PIL.Image: Blurred PIL Image object
+    """
+    try:
+        import requests
+        from PIL import Image, ImageFilter
+        import io
+        
+        # Download image
+        response = requests.get(image_url, timeout=30)
+        if response.status_code != 200:
+            return None
+        
+        # Convert to PIL Image
+        image_data = io.BytesIO(response.content)
+        image = Image.open(image_data)
+        
+        # Apply heavy Gaussian blur
+        blurred_image = image.filter(ImageFilter.GaussianBlur(radius=blur_strength))
+        
+        return blurred_image
+        
+    except Exception as e:
+        print(f"⚠️ Failed to apply Gaussian blur: {e}")
+        return None
+
+# Initialize NSFW detector (download from Hugging Face)
 try:
-    # Load compiled modules
-    util_module = load_pyc_module("util", lib_dir / "util.pyc")
-    nfsw_module = load_pyc_module("nfsw", lib_dir / "nfsw.pyc")
+    nsfw_detector = NSFWDetector()  # Auto download falconsai_yolov9_nsfw_model_quantized.pt from Hugging Face
+    print("✅ NSFW detector initialized successfully")
+except Exception as e:
+    print(f"❌ NSFW detector initialization failed: {e}")
+    nsfw_detector = None
+
+def edit_image_interface(input_image, prompt, request: gr.Request, progress=gr.Progress()):
+    """
+    Interface function for processing image editing with phase-based limitations
+    """
+    try:
+        # Extract user IP
+        client_ip = request.client.host
+        x_forwarded_for = dict(request.headers).get('x-forwarded-for')
+        if x_forwarded_for:
+            client_ip = x_forwarded_for   
+        if client_ip not in IP_Dict:
+            IP_Dict[client_ip] = 0
+        IP_Dict[client_ip] += 1
+
+        if input_image is None:
+            return None, "Please upload an image first", gr.update(visible=False)
+        
+        if not prompt or prompt.strip() == "":
+            return None, "Please enter editing prompt", gr.update(visible=False)
+        
+        # Check if prompt length is greater than 3 characters
+        if len(prompt.strip()) <= 3:
+            return None, "❌ Editing prompt must be more than 3 characters", gr.update(visible=False)
+    except Exception as e:
+        print(f"⚠️ Request preprocessing error: {e}")
+        return None, "❌ Request processing error", gr.update(visible=False)
+    
+    # Get user current phase
+    current_phase = get_ip_phase(client_ip)
+    current_count = get_ip_generation_count(client_ip)
+    
+    print(f"📊 User phase info - IP: {client_ip}, current phase: {current_phase}, generation count: {current_count}")
+    
+    # Check if user reached the like button tip threshold  
+    show_like_tip = (current_count >= TIP_TRY_N)
+    
+    # Check if completely blocked
+    if current_phase == 'blocked':
+        # Generate blocked limit button
+        blocked_button_html = f"""
+        <div style='display: flex; justify-content: center; gap: 15px; margin: 10px 0 5px 0; padding: 0px;'>
+            <a href='https://omnicreator.net/#generator' target='_blank' style='
+                display: inline-flex; 
+                align-items: center;
+                justify-content: center;
+                padding: 16px 32px; 
+                background: linear-gradient(135deg, #e74c3c 0%, #c0392b 100%);
+                color: white; 
+                text-decoration: none; 
+                border-radius: 12px; 
+                font-weight: 600;
+                font-size: 16px;
+                text-align: center;
+                min-width: 200px;
+                box-shadow: 0 4px 15px rgba(231, 76, 60, 0.4);
+                transition: all 0.3s ease;
+                border: none;
+            '>&#128640; Unlimited Generation</a>
+        </div>
+        """
+        return None, f"❌ You have reached Hugging Face's free generation limit. Please visit https://omnicreator.net/#generator for unlimited generation", gr.update(value=blocked_button_html, visible=True)
+    
+    # Check rate limit (applies to rate_limit phases)
+    if current_phase in ['rate_limit_1', 'rate_limit_2', 'rate_limit_3']:
+        is_limited, wait_minutes, window_count = check_rate_limit_for_phase(client_ip, current_phase)
+        if is_limited:
+            wait_minutes_int = int(wait_minutes) + 1
+            # Generate rate limit button
+            rate_limit_button_html = f"""
+            <div style='display: flex; justify-content: center; gap: 15px; margin: 10px 0 5px 0; padding: 0px;'>
+                <a href='https://omnicreator.net/#generator' target='_blank' style='
+                    display: inline-flex; 
+                    align-items: center;
+                    justify-content: center;
+                    padding: 16px 32px; 
+                    background: linear-gradient(135deg, #f39c12 0%, #e67e22 100%);
+                    color: white; 
+                    text-decoration: none; 
+                    border-radius: 12px; 
+                    font-weight: 600;
+                    font-size: 16px;
+                    text-align: center;
+                    min-width: 200px;
+                    box-shadow: 0 4px 15px rgba(243, 156, 18, 0.4);
+                    transition: all 0.3s ease;
+                    border: none;
+                '>⏰ Skip Wait - Unlimited Generation</a>
+            </div>
+            """
+            return None, f"❌ You have reached Hugging Face's free generation limit. Please visit https://omnicreator.net/#generator for unlimited generation, or wait {wait_minutes_int} minutes before generating again", gr.update(value=rate_limit_button_html, visible=True)
+    
+    # Handle NSFW detection based on phase
+    is_nsfw_task = False  # Track if this task involves NSFW content
+    
+    # Skip NSFW detection in free phase
+    if current_phase != 'free' and nsfw_detector is not None and input_image is not None:
+        try:
+            nsfw_result = nsfw_detector.predict_pil_label_only(input_image)
+            
+            if nsfw_result.lower() == "nsfw":
+                is_nsfw_task = True
+                print(f"🔍 Input NSFW detected in {current_phase} phase: ❌❌❌ {nsfw_result} - IP: {client_ip} (will blur result)")
+            else:
+                print(f"🔍 Input NSFW check passed: ✅✅✅ {nsfw_result} - IP: {client_ip}")
+                
+        except Exception as e:
+            print(f"⚠️ Input NSFW detection failed: {e}")
+            # Allow continuation when detection fails
+    
+    result_url = None
+    status_message = ""
+    
+    def progress_callback(message):
+        try:
+            nonlocal status_message
+            status_message = message
+            # Add error handling to prevent progress update failure
+            if progress is not None:
+                # Enhanced progress display with better formatting
+                if "Queue:" in message or "tasks ahead" in message:
+                    # Queue status - show with different progress value to indicate waiting
+                    progress(0.1, desc=message)
+                elif "Processing" in message or "AI is processing" in message:
+                    # Processing status
+                    progress(0.7, desc=message)
+                elif "Generating" in message or "Almost done" in message:
+                    # Generation status
+                    progress(0.9, desc=message)
+                else:
+                    # Default status
+                    progress(0.5, desc=message)
+        except Exception as e:
+            print(f"⚠️ Progress update failed: {e}")
+    
+    try:
+        # Record generation attempt (before actual generation to ensure correct count)
+        record_generation_attempt(client_ip, current_phase)
+        updated_count = get_ip_generation_count(client_ip)
+        
+        print(f"✅ Processing started - IP: {client_ip}, phase: {current_phase}, total count: {updated_count}, prompt: {prompt.strip()}", flush=True)
+        
+        # Call image editing processing function
+        result_url, message, task_uuid = process_image_edit(input_image, prompt.strip(), None, progress_callback)
+        
+        if result_url:
+            print(f"✅ Processing completed successfully - IP: {client_ip}, result_url: {result_url}, task_uuid: {task_uuid}", flush=True)
+            
+            # Detect result image NSFW content (only in rate limit phases)
+            if nsfw_detector is not None and current_phase != 'free':
+                try:
+                    if progress is not None:
+                        progress(0.9, desc="Checking result image...")
+                    
+                    is_nsfw, nsfw_error = download_and_check_result_nsfw(result_url, nsfw_detector)
+                    
+                    if nsfw_error:
+                        print(f"⚠️ Result image NSFW detection error - IP: {client_ip}, error: {nsfw_error}")
+                    elif is_nsfw:
+                        is_nsfw_task = True  # Mark task as NSFW
+                        print(f"🔍 Result image NSFW detected in {current_phase} phase: ❌❌❌ - IP: {client_ip} (will blur result)")
+                    else:
+                        print(f"🔍 Result image NSFW check passed: ✅✅✅ - IP: {client_ip}")
+                        
+                except Exception as e:
+                    print(f"��️ Result image NSFW detection exception - IP: {client_ip}, error: {str(e)}")
+            
+            # Apply blur if this is an NSFW task in rate limit phases
+            should_blur = False
+            
+            if current_phase in ['rate_limit_1', 'rate_limit_2', 'rate_limit_3'] and is_nsfw_task:
+                should_blur = True
+            
+            # Apply blur processing
+            if should_blur:
+                if progress is not None:
+                    progress(0.95, desc="Applying content filter...")
+                
+                blurred_image = apply_gaussian_blur_to_image_url(result_url)
+                if blurred_image is not None:
+                    final_result = blurred_image  # Return PIL Image object
+                    final_message = f"⚠️ NSFW content detected, content filter applied. NSFW content is prohibited by Hugging Face, but you can generate unlimited content at our official website https://omnicreator.net/#generator"
+                    print(f"🔒 Applied Gaussian blur for NSFW content - IP: {client_ip}")
+                else:
+                    # Blur failed, return original URL with warning
+                    final_result = result_url
+                    final_message = f"⚠️ NSFW content detected, but content filter failed. Please visit https://omnicreator.net/#generator for better experience"
+                
+                # Generate NSFW button for blurred content
+                nsfw_action_buttons_html = f"""
+                <div style='display: flex; justify-content: center; gap: 15px; margin: 10px 0 5px 0; padding: 0px;'>
+                    <a href='https://omnicreator.net/#generator' target='_blank' style='
+                        display: inline-flex; 
+                        align-items: center;
+                        justify-content: center;
+                        padding: 16px 32px; 
+                        background: linear-gradient(135deg, #ff6b6b 0%, #feca57 100%);
+                        color: white; 
+                        text-decoration: none; 
+                        border-radius: 12px; 
+                        font-weight: 600;
+                        font-size: 16px;
+                        text-align: center;
+                        min-width: 200px;
+                        box-shadow: 0 4px 15px rgba(255, 107, 107, 0.4);
+                        transition: all 0.3s ease;
+                        border: none;
+                    '>🔥 Unlimited NSFW Generation</a>
+                </div>
+                """
+                return final_result, final_message, gr.update(value=nsfw_action_buttons_html, visible=True)
+            else:
+                final_result = result_url
+                final_message = "✅ " + message
+            
+            try:
+                if progress is not None:
+                    progress(1.0, desc="Processing completed")
+            except Exception as e:
+                print(f"⚠️ Final progress update failed: {e}")
+            
+            # Generate action buttons HTML like Trump AI Voice
+            action_buttons_html = ""
+            if task_uuid:
+                task_detail_url = f"https://omnicreator.net/my-creations/task/{task_uuid}"
+                action_buttons_html = f"""
+                <div style='display: flex; justify-content: center; gap: 15px; margin: 10px 0 5px 0; padding: 0px;'>
+                    <a href='{task_detail_url}' target='_blank' style='
+                        display: inline-flex; 
+                        align-items: center;
+                        justify-content: center;
+                        padding: 16px 32px; 
+                        background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
+                        color: white; 
+                        text-decoration: none; 
+                        border-radius: 12px; 
+                        font-weight: 600;
+                        font-size: 16px;
+                        text-align: center;
+                        min-width: 160px;
+                        box-shadow: 0 4px 15px rgba(102, 126, 234, 0.4);
+                        transition: all 0.3s ease;
+                        border: none;
+                    '>&#128444; Download HD Image</a>
+                    <a href='https://omnicreator.net/#generator' target='_blank' style='
+                        display: inline-flex; 
+                        align-items: center;
+                        justify-content: center;
+                        padding: 16px 32px; 
+                        background: linear-gradient(135deg, #11998e 0%, #38ef7d 100%);
+                        color: white; 
+                        text-decoration: none; 
+                        border-radius: 12px; 
+                        font-weight: 600;
+                        font-size: 16px;
+                        text-align: center;
+                        min-width: 160px;
+                        box-shadow: 0 4px 15px rgba(17, 153, 142, 0.4);
+                        transition: all 0.3s ease;
+                        border: none;
+                    '>&#128640; Unlimited Generation</a>
+                </div>
+                """
+            
+            # Add popup script if needed (using different approach)
+            if show_like_tip:
+                action_buttons_html += """
+                <div style='display: flex; justify-content: center; margin: 15px 0 5px 0; padding: 0px;'>
+                    <div style='
+                        display: inline-flex; 
+                        align-items: center;
+                        justify-content: center;
+                        padding: 12px 24px; 
+                        background: linear-gradient(135deg, #ff6b6b 0%, #feca57 100%);
+                        color: white; 
+                        border-radius: 10px; 
+                        font-weight: 600;
+                        font-size: 14px;
+                        text-align: center;
+                        max-width: 400px;
+                        box-shadow: 0 3px 12px rgba(255, 107, 107, 0.3);
+                        border: none;
+                    '>👉 Click the ❤️ Like button to unlock more free trial attempts!</div>
+                </div>
+                """
+            
+            return final_result, final_message, gr.update(value=action_buttons_html, visible=True)
+        else:
+            print(f"❌ Processing failed - IP: {client_ip}, error: {message}", flush=True)
+            return None, "❌ " + message, gr.update(visible=False)
+            
+    except Exception as e:
+        print(f"❌ Processing exception - IP: {client_ip}, error: {str(e)}")
+        return None, f"❌ Error occurred during processing: {str(e)}", gr.update(visible=False)
+
+def local_edit_interface(image_dict, prompt, reference_image, request: gr.Request, progress=gr.Progress()):
+    """
+    Handle local editing requests (with phase-based limitations)
+    """
+    try:
+        # Extract user IP
+        client_ip = request.client.host
+        x_forwarded_for = dict(request.headers).get('x-forwarded-for')
+        if x_forwarded_for:
+            client_ip = x_forwarded_for   
+        if client_ip not in IP_Dict:
+            IP_Dict[client_ip] = 0
+        IP_Dict[client_ip] += 1
+
+        if image_dict is None:
+            return None, "Please upload an image and draw the area to edit", gr.update(visible=False)
+        
+        # Handle different input formats for ImageEditor
+        if isinstance(image_dict, dict):
+            # ImageEditor dict format
+            if "background" not in image_dict or "layers" not in image_dict:
+                return None, "Please draw the area to edit on the image", gr.update(visible=False)
+            
+            base_image = image_dict["background"]
+            layers = image_dict["layers"]
+            
+            # Special handling: if background is None but composite exists, use composite
+            if base_image is None and "composite" in image_dict and image_dict["composite"] is not None:
+                print("🔧 Background is None, using composite instead")
+                base_image = image_dict["composite"]
+        else:
+            # Simple case: Direct PIL Image (from example)
+            base_image = image_dict
+            layers = []
+        
+        # Check for special example case - bypass mask requirement
+        is_example_case = prompt and prompt.startswith("EXAMPLE_PANDA_CAT_")
+        
+        # Debug: check current state
+        if is_example_case:
+            print(f"🔍 Example case detected - base_image is None: {base_image is None}")
+        
+        # Special handling for example case: load image directly from file
+        if is_example_case and base_image is None:
+            try:
+                from PIL import Image
+                import os
+                
+                main_path = "datas/panda01.jpeg"
+                print(f"🔍 Trying to load: {main_path}, exists: {os.path.exists(main_path)}")
+                
+                if os.path.exists(main_path):
+                    base_image = Image.open(main_path)
+                    print(f"✅ Successfully loaded example image: {base_image.size}")
+                else:
+                    return None, f"❌ Example image not found: {main_path}", gr.update(visible=False)
+            except Exception as e:
+                return None, f"❌ Failed to load example image: {str(e)}", gr.update(visible=False)
+        
+        # Additional check for base_image
+        if base_image is None:
+            if is_example_case:
+                print(f"❌ Example case but base_image still None!")
+            return None, "❌ No image found. Please upload an image first.", gr.update(visible=False)
+        
+        if not layers and not is_example_case:
+            return None, "Please draw the area to edit on the image", gr.update(visible=False)
+            
+        if not prompt or prompt.strip() == "":
+            return None, "Please enter editing prompt", gr.update(visible=False)
+        
+        # Check prompt length
+        if len(prompt.strip()) <= 3:
+            return None, "❌ Editing prompt must be more than 3 characters", gr.update(visible=False)
+    except Exception as e:
+        print(f"⚠️ Local edit request preprocessing error: {e}")
+        return None, "❌ Request processing error", gr.update(visible=False)
+    
+    # Get user current phase
+    current_phase = get_ip_phase(client_ip)
+    current_count = get_ip_generation_count(client_ip)
+    
+    print(f"📊 Local edit user phase info - IP: {client_ip}, current phase: {current_phase}, generation count: {current_count}")
+    
+    # Check if user reached the like button tip threshold  
+    show_like_tip = (current_count >= TIP_TRY_N)
     
-    # Import app module (source file)
-    import app as app_module
+    # Check if completely blocked
+    if current_phase == 'blocked':
+        # Generate blocked limit button
+        blocked_button_html = f"""
+        <div style='display: flex; justify-content: center; gap: 15px; margin: 10px 0 5px 0; padding: 0px;'>
+            <a href='https://omnicreator.net/#generator' target='_blank' style='
+                display: inline-flex; 
+                align-items: center;
+                justify-content: center;
+                padding: 16px 32px; 
+                background: linear-gradient(135deg, #e74c3c 0%, #c0392b 100%);
+                color: white; 
+                text-decoration: none; 
+                border-radius: 12px; 
+                font-weight: 600;
+                font-size: 16px;
+                text-align: center;
+                min-width: 200px;
+                box-shadow: 0 4px 15px rgba(231, 76, 60, 0.4);
+                transition: all 0.3s ease;
+                border: none;
+            '>&#128640; Unlimited Generation</a>
+        </div>
+        """
+        return None, f"❌ You have reached Hugging Face's free generation limit. Please visit https://omnicreator.net/#generator for unlimited generation", gr.update(value=blocked_button_html, visible=True)
+    
+    # Check rate limit (applies to rate_limit phases)
+    if current_phase in ['rate_limit_1', 'rate_limit_2', 'rate_limit_3']:
+        is_limited, wait_minutes, window_count = check_rate_limit_for_phase(client_ip, current_phase)
+        if is_limited:
+            wait_minutes_int = int(wait_minutes) + 1
+            # Generate rate limit button
+            rate_limit_button_html = f"""
+            <div style='display: flex; justify-content: center; gap: 15px; margin: 10px 0 5px 0; padding: 0px;'>
+                <a href='https://omnicreator.net/#generator' target='_blank' style='
+                    display: inline-flex; 
+                    align-items: center;
+                    justify-content: center;
+                    padding: 16px 32px; 
+                    background: linear-gradient(135deg, #f39c12 0%, #e67e22 100%);
+                    color: white; 
+                    text-decoration: none; 
+                    border-radius: 12px; 
+                    font-weight: 600;
+                    font-size: 16px;
+                    text-align: center;
+                    min-width: 200px;
+                    box-shadow: 0 4px 15px rgba(243, 156, 18, 0.4);
+                    transition: all 0.3s ease;
+                    border: none;
+                '>⏰ Skip Wait - Unlimited Generation</a>
+            </div>
+            """
+            return None, f"❌ You have reached Hugging Face's free generation limit. Please visit https://omnicreator.net/#generator for unlimited generation, or wait {wait_minutes_int} minutes before generating again", gr.update(value=rate_limit_button_html, visible=True)
+    
+    # Handle NSFW detection based on phase
+    is_nsfw_task = False  # Track if this task involves NSFW content
+    
+    # Skip NSFW detection in free phase
+    if current_phase != 'free' and nsfw_detector is not None and base_image is not None:
+        try:
+            nsfw_result = nsfw_detector.predict_pil_label_only(base_image)
+            
+            if nsfw_result.lower() == "nsfw":
+                is_nsfw_task = True
+                print(f"🔍 Local edit input NSFW detected in {current_phase} phase: ❌❌❌ {nsfw_result} - IP: {client_ip} (will blur result)")
+            else:
+                print(f"🔍 Local edit input NSFW check passed: ✅✅✅ {nsfw_result} - IP: {client_ip}")
+                
+        except Exception as e:
+            print(f"⚠️ Local edit input NSFW detection failed: {e}")
+            # Allow continuation when detection fails
+
+    result_url = None
+    status_message = ""
+    
+    def progress_callback(message):
+        try:
+            nonlocal status_message
+            status_message = message
+            # Add error handling to prevent progress update failure
+            if progress is not None:
+                # Enhanced progress display with better formatting for local editing
+                if "Queue:" in message or "tasks ahead" in message:
+                    # Queue status - show with different progress value to indicate waiting
+                    progress(0.1, desc=message)
+                elif "Processing" in message or "AI is processing" in message:
+                    # Processing status
+                    progress(0.7, desc=message)
+                elif "Generating" in message or "Almost done" in message:
+                    # Generation status
+                    progress(0.9, desc=message)
+                else:
+                    # Default status
+                    progress(0.5, desc=message)
+        except Exception as e:
+            print(f"⚠️ Local edit progress update failed: {e}")
+    
+    try:
+        # Record generation attempt (before actual generation to ensure correct count)
+        record_generation_attempt(client_ip, current_phase)
+        updated_count = get_ip_generation_count(client_ip)
+        
+        print(f"✅ Local editing started - IP: {client_ip}, phase: {current_phase}, total count: {updated_count}, prompt: {prompt.strip()}", flush=True)
+        
+        # Clean prompt for API call
+        clean_prompt = prompt.strip()
+        if clean_prompt.startswith("EXAMPLE_PANDA_CAT_"):
+            clean_prompt = clean_prompt[18:]  # Remove the prefix
+        
+        # Call local image editing processing function
+        if is_example_case:
+            # For example case, pass special flag to use local mask file
+            result_url, message, task_uuid = process_local_image_edit(base_image, layers, clean_prompt, reference_image, progress_callback, use_example_mask="datas/panda01m.jpeg")
+        else:
+            # Normal case
+            result_url, message, task_uuid = process_local_image_edit(base_image, layers, clean_prompt, reference_image, progress_callback)
+        
+        if result_url:
+            print(f"✅ Local editing completed successfully - IP: {client_ip}, result_url: {result_url}, task_uuid: {task_uuid}", flush=True)
+            
+            # Detect result image NSFW content (only in rate limit phases)
+            if nsfw_detector is not None and current_phase != 'free':
+                try:
+                    if progress is not None:
+                        progress(0.9, desc="Checking result image...")
+                    
+                    is_nsfw, nsfw_error = download_and_check_result_nsfw(result_url, nsfw_detector)
+                    
+                    if nsfw_error:
+                        print(f"⚠️ Local edit result image NSFW detection error - IP: {client_ip}, error: {nsfw_error}")
+                    elif is_nsfw:
+                        is_nsfw_task = True  # Mark task as NSFW
+                        print(f"🔍 Local edit result image NSFW detected in {current_phase} phase: ❌❌❌ - IP: {client_ip} (will blur result)")
+                    else:
+                        print(f"🔍 Local edit result image NSFW check passed: ✅✅✅ - IP: {client_ip}")
+                        
+                except Exception as e:
+                    print(f"⚠️ Local edit result image NSFW detection exception - IP: {client_ip}, error: {str(e)}")
+            
+            # Apply blur if this is an NSFW task in rate limit phases
+            should_blur = False
+            
+            if current_phase in ['rate_limit_1', 'rate_limit_2', 'rate_limit_3'] and is_nsfw_task:
+                should_blur = True
+            
+            # Apply blur processing
+            if should_blur:
+                if progress is not None:
+                    progress(0.95, desc="Applying content filter...")
+                
+                blurred_image = apply_gaussian_blur_to_image_url(result_url)
+                if blurred_image is not None:
+                    final_result = blurred_image  # Return PIL Image object
+                    final_message = f"⚠️ NSFW content detected, content filter applied. NSFW content is prohibited by Hugging Face, but you can generate unlimited content at our official website https://omnicreator.net/#generator"
+                    print(f"🔒 Local edit applied Gaussian blur for NSFW content - IP: {client_ip}")
+                else:
+                    # Blur failed, return original URL with warning
+                    final_result = result_url
+                    final_message = f"⚠️ NSFW content detected, but content filter failed. Please visit https://omnicreator.net/#generator for better experience"
+                
+                # Generate NSFW button for blurred content
+                nsfw_action_buttons_html = f"""
+                <div style='display: flex; justify-content: center; gap: 15px; margin: 10px 0 5px 0; padding: 0px;'>
+                    <a href='https://omnicreator.net/#generator' target='_blank' style='
+                        display: inline-flex; 
+                        align-items: center;
+                        justify-content: center;
+                        padding: 16px 32px; 
+                        background: linear-gradient(135deg, #ff6b6b 0%, #feca57 100%);
+                        color: white; 
+                        text-decoration: none; 
+                        border-radius: 12px; 
+                        font-weight: 600;
+                        font-size: 16px;
+                        text-align: center;
+                        min-width: 200px;
+                        box-shadow: 0 4px 15px rgba(255, 107, 107, 0.4);
+                        transition: all 0.3s ease;
+                        border: none;
+                    '>🔥 Unlimited NSFW Generation</a>
+                </div>
+                """
+                return final_result, final_message, gr.update(value=nsfw_action_buttons_html, visible=True)
+            else:
+                final_result = result_url
+                final_message = "✅ " + message
+            
+            try:
+                if progress is not None:
+                    progress(1.0, desc="Processing completed")
+            except Exception as e:
+                print(f"⚠️ Local edit final progress update failed: {e}")
+            
+            # Generate action buttons HTML like Trump AI Voice
+            action_buttons_html = ""
+            if task_uuid:
+                task_detail_url = f"https://omnicreator.net/my-creations/task/{task_uuid}"
+                action_buttons_html = f"""
+                <div style='display: flex; justify-content: center; gap: 15px; margin: 10px 0 5px 0; padding: 0px;'>
+                    <a href='{task_detail_url}' target='_blank' style='
+                        display: inline-flex; 
+                        align-items: center;
+                        justify-content: center;
+                        padding: 16px 32px; 
+                        background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
+                        color: white; 
+                        text-decoration: none; 
+                        border-radius: 12px; 
+                        font-weight: 600;
+                        font-size: 16px;
+                        text-align: center;
+                        min-width: 160px;
+                        box-shadow: 0 4px 15px rgba(102, 126, 234, 0.4);
+                        transition: all 0.3s ease;
+                        border: none;
+                    '>&#128444; Download HD Image</a>
+                    <a href='https://omnicreator.net/#generator' target='_blank' style='
+                        display: inline-flex; 
+                        align-items: center;
+                        justify-content: center;
+                        padding: 16px 32px; 
+                        background: linear-gradient(135deg, #11998e 0%, #38ef7d 100%);
+                        color: white; 
+                        text-decoration: none; 
+                        border-radius: 12px; 
+                        font-weight: 600;
+                        font-size: 16px;
+                        text-align: center;
+                        min-width: 160px;
+                        box-shadow: 0 4px 15px rgba(17, 153, 142, 0.4);
+                        transition: all 0.3s ease;
+                        border: none;
+                    '>&#128640; Unlimited Generation</a>
+                </div>
+                """
+            
+            # Add popup script if needed (using different approach)
+            if show_like_tip:
+                action_buttons_html += """
+                <div style='display: flex; justify-content: center; margin: 15px 0 5px 0; padding: 0px;'>
+                    <div style='
+                        display: inline-flex; 
+                        align-items: center;
+                        justify-content: center;
+                        padding: 12px 24px; 
+                        background: linear-gradient(135deg, #ff6b6b 0%, #feca57 100%);
+                        color: white; 
+                        border-radius: 10px; 
+                        font-weight: 600;
+                        font-size: 14px;
+                        text-align: center;
+                        max-width: 400px;
+                        box-shadow: 0 3px 12px rgba(255, 107, 107, 0.3);
+                        border: none;
+                    '>👉 Please consider clicking the ❤️ Like button to support this space!</div>
+                </div>
+                """
+            
+            return final_result, final_message, gr.update(value=action_buttons_html, visible=True)
+        else:
+            print(f"❌ Local editing processing failed - IP: {client_ip}, error: {message}", flush=True)
+            return None, "❌ " + message, gr.update(visible=False)
+            
+    except Exception as e:
+        print(f"❌ Local editing exception - IP: {client_ip}, error: {str(e)}")
+        return None, f"❌ Error occurred during processing: {str(e)}", gr.update(visible=False)
+
+# Create Gradio interface
+def create_app():
+    with gr.Blocks(
+        title="AI Image Editor",
+        theme=gr.themes.Soft(),
+        css="""
+        .main-container {
+            max-width: 1200px;
+            margin: 0 auto;
+        }
+        .upload-area {
+            border: 2px dashed #ccc;
+            border-radius: 10px;
+            padding: 20px;
+            text-align: center;
+        }
+        .result-area {
+            margin-top: 20px;
+            padding: 20px;
+            border-radius: 10px;
+            background-color: #f8f9fa;
+        }
+        .use-as-input-btn {
+            margin-top: 10px;
+            width: 100%;
+        }
+        """,
+        # Improve concurrency performance configuration
+        head="""
+        <script>
+        // Reduce client-side state update frequency, avoid excessive SSE connections
+        if (window.gradio) {
+            window.gradio.update_frequency = 2000; // Update every 2 seconds
+        }
+        </script>
+        """
+    ) as app:
+        
+        # Main title - styled like Trump AI Voice
+        gr.HTML("""
+        <div style="text-align: center; margin: 5px auto 0px auto; max-width: 800px;">
+            <h1 style="color: #2c3e50; margin: 0; font-size: 3.5em; font-weight: 800; letter-spacing: 3px; text-shadow: 2px 2px 4px rgba(0,0,0,0.1);">
+                🎨 AI Image Editor
+            </h1>
+        </div>
+        """, padding=False)
+        
+        # 🌟 NEW: Multi-Image Editing Announcement Banner with breathing effect
+        gr.HTML("""
+        <style>
+        @keyframes breathe {
+            0%, 100% { transform: scale(1); }
+            50% { transform: scale(1.02); }
+        }
+        .breathing-banner {
+            animation: breathe 3s ease-in-out infinite;
+        }
+        </style>
+        <div class="breathing-banner" style="
+            background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
+            margin: 5px auto 5px auto;
+            padding: 6px 40px;
+            border-radius: 20px;
+            max-width: 700px;
+            box-shadow: 0 2px 8px rgba(102, 126, 234, 0.3);
+            text-align: center;
+        ">
+            <span style="color: white; font-weight: 600; font-size: 1.0em;">
+                🚀 NEWS: 
+                <a href="https://huggingface.co/spaces/Selfit/Multi-Image-Edit" target="_blank" style="
+                    color: white; 
+                    text-decoration: none; 
+                    border-bottom: 1px solid rgba(255,255,255,0.5);
+                    transition: all 0.3s ease;
+                " onmouseover="this.style.borderBottom='1px solid white'" 
+                   onmouseout="this.style.borderBottom='1px solid rgba(255,255,255,0.5)'">
+                   World's First Multi-Image Editing Tool →
+                </a>
+            </span>
+        </div>
+        """, padding=False)
+        
+        with gr.Tabs():
+            with gr.Tab("🌍 Global Editor"):
+                with gr.Row():
+                    with gr.Column(scale=1):
+                        gr.Markdown("### 📸 Upload Image")
+                        input_image = gr.Image(
+                            label="Select image to edit",
+                            type="pil",
+                            height=512,
+                            elem_classes=["upload-area"]
+                        )
+
+                        gr.Markdown("### ✍️ Editing Instructions")
+                        prompt_input = gr.Textbox(
+                            label="Enter editing prompt",
+                            placeholder="For example: change background to beach, add rainbow, remove background, etc...",
+                            lines=3,
+                            max_lines=5
+                        )
+                        
+                        edit_button = gr.Button(
+                            "🚀 Start Editing",
+                            variant="primary",
+                            size="lg"
+                        )
+                    
+                    with gr.Column(scale=1):
+                        gr.Markdown("### 🎯 Editing Result")
+                        output_image = gr.Image(
+                            label="Edited image",
+                            height=320,
+                            elem_classes=["result-area"]
+                        )
+                        
+                        use_as_input_btn = gr.Button(
+                            "🔄 Use as Input",
+                            variant="secondary",
+                            size="sm",
+                            elem_classes=["use-as-input-btn"]
+                        )
+                        
+                        status_output = gr.Textbox(
+                            label="Processing status",
+                            lines=2,
+                            max_lines=3,
+                            interactive=False
+                        )
+                        
+                        action_buttons = gr.HTML(visible=False)
+                
+                gr.Markdown("### 💡 Prompt Examples")
+                with gr.Row():
+                    example_prompts = [
+                        "Set the background to a grand opera stage with red curtains",
+                        "Change the outfit into a traditional Chinese hanfu with flowing sleeves",
+                        "Give the character blue dragon-like eyes with glowing pupils",
+                        "Change lighting to soft dreamy pastel glow",
+                        "Change pose to sitting cross-legged on the ground"
+                    ]
+                    
+                    for prompt in example_prompts:
+                        gr.Button(
+                            prompt,
+                            size="sm"
+                        ).click(
+                            lambda p=prompt: p,
+                            outputs=prompt_input
+                        )
+                
+                edit_button.click(
+                    fn=edit_image_interface,
+                    inputs=[input_image, prompt_input],
+                    outputs=[output_image, status_output, action_buttons],
+                    show_progress=True,
+                    concurrency_limit=10,
+                    api_name="global_edit"
+                )
+                
+                def simple_use_as_input(output_img):
+                    if output_img is not None:
+                        return output_img
+                    return None
+                
+                use_as_input_btn.click(
+                    fn=simple_use_as_input,
+                    inputs=[output_image],
+                    outputs=[input_image]
+                )
+
+            with gr.Tab("🖌️ Local Inpaint"):
+                with gr.Row():
+                    with gr.Column(scale=1):
+                        gr.Markdown("### 📸 Upload Image and Draw Mask")
+                        local_input_image = gr.ImageEditor(
+                            label="Upload image and draw mask",
+                            type="pil",
+                            height=512,
+                            brush=gr.Brush(colors=["#ff0000"], default_size=180),
+                            elem_classes=["upload-area"]
+                        )
+
+                        gr.Markdown("### 🖼️ Reference Image(Optional)")
+                        local_reference_image = gr.Image(
+                            label="Upload reference image (optional)",
+                            type="pil",
+                            height=256
+                        )
+                        
+                        gr.Markdown("### ✍️ Editing Instructions")
+                        local_prompt_input = gr.Textbox(
+                            label="Enter local editing prompt",
+                            placeholder="For example: change selected area hair to golden, add patterns to selected object, change selected area color, etc...",
+                            lines=3,
+                            max_lines=5
+                        )
+                        
+                        local_edit_button = gr.Button(
+                            "🎯 Start Local Editing",
+                            variant="primary",
+                            size="lg"
+                        )
+                    
+                    with gr.Column(scale=1):
+                        gr.Markdown("### 🎯 Editing Result")
+                        local_output_image = gr.Image(
+                            label="Local edited image",
+                            height=320,
+                            elem_classes=["result-area"]
+                        )
+                        
+                        local_use_as_input_btn = gr.Button(
+                            "🔄 Use as Input",
+                            variant="secondary",
+                            size="sm",
+                            elem_classes=["use-as-input-btn"]
+                        )
+                        
+                        local_status_output = gr.Textbox(
+                            label="Processing status",
+                            lines=2,
+                            max_lines=3,
+                            interactive=False
+                        )
+                        
+                        local_action_buttons = gr.HTML(visible=False)
+             
+                local_edit_button.click(
+                    fn=local_edit_interface,
+                    inputs=[local_input_image, local_prompt_input, local_reference_image],
+                    outputs=[local_output_image, local_status_output, local_action_buttons],
+                    show_progress=True,
+                    concurrency_limit=8,
+                    api_name="local_edit"
+                )
+                
+                def simple_local_use_as_input(output_img):
+                    if output_img is not None:
+                        return {
+                            "background": output_img,
+                            "layers": [],
+                            "composite": output_img
+                        }
+                    return None
+                
+                local_use_as_input_btn.click(
+                    fn=simple_local_use_as_input,
+                    inputs=[local_output_image],
+                    outputs=[local_input_image]
+                )
+                
+                # Local inpaint example
+                gr.Markdown("### 💡 Local Inpaint Example")
+                
+                def load_local_example():
+                    """Load panda to cat transformation example - simplified, mask handled in backend"""
+                    try:
+                        from PIL import Image
+                        import os
+                        
+                        # Check file paths
+                        main_path = "datas/panda01.jpeg"
+                        ref_path = "datas/cat01.webp"
+                        
+                        # Load main image
+                        if not os.path.exists(main_path):
+                            return None, None, "EXAMPLE_PANDA_CAT_let the cat ride on the panda"
+                        
+                        main_img = Image.open(main_path)
+                        
+                        # Load reference image  
+                        if not os.path.exists(ref_path):
+                            ref_img = None
+                        else:
+                            ref_img = Image.open(ref_path)
+                        
+                        # ImageEditor format
+                        editor_data = {
+                            "background": main_img,
+                            "layers": [],
+                            "composite": main_img
+                        }
+                        
+                        # Special prompt to indicate this is the example case
+                        prompt = "EXAMPLE_PANDA_CAT_let the cat ride on the panda"
+                        
+                        # Return just the PIL image instead of dict format to avoid UI state issues
+                        return main_img, ref_img, prompt
+                        
+                    except Exception as e:
+                        return None, None, "EXAMPLE_PANDA_CAT_Transform the panda head into a cute cat head, keeping the body"
+                    
+                    # Example display
+                gr.Markdown("#### 🐼➡️🐱 Example: Panda to Cat Transformation")
+                with gr.Row():
+                    with gr.Column(scale=2):
+                        # Preview images for local example
+                        with gr.Row():
+                            try:
+                                gr.Image("datas/panda01.jpeg", label="Main Image", height=120, width=120, show_label=True, interactive=False)
+                                gr.Image("datas/panda01m.jpeg", label="Mask", height=120, width=120, show_label=True, interactive=False)
+                                gr.Image("datas/cat01.webp", label="Reference", height=120, width=120, show_label=True, interactive=False)
+                            except:
+                                gr.Markdown("*Preview images not available*")
+                        gr.Markdown("**Prompt**: let the cat ride on the panda  \n**Note**: Mask will be automatically applied when you submit this example")
+                    with gr.Column(scale=1):
+                        gr.Button(
+                            "🎨 Load Panda Example",
+                            size="lg",
+                            variant="secondary"
+                        ).click(
+                            fn=load_local_example,
+                            outputs=[local_input_image, local_reference_image, local_prompt_input]
+                        )
+                
+                # Add a refresh button to fix UI state issues
+                gr.Button(
+                    "🔄 Refresh Image Editor", 
+                    size="sm",
+                    variant="secondary"
+                ).click(
+                    fn=lambda: gr.update(),
+                    outputs=[local_input_image]
+                )
+        
+        # SEO Content Section
+        gr.HTML("""
+    <div style="width: 100%; margin: 50px 0; padding: 0 20px;">
+        
+        <div style="text-align: center; background: linear-gradient(135deg, #667eea 0%, #764ba2 100%); color: white; padding: 40px; border-radius: 20px; margin: 40px 0;">
+            <h2 style="margin: 0 0 20px 0; font-size: 2.2em; font-weight: 700;">
+                &#127912; Unlimited AI Image Generation & Editing
+            </h2>
+            <p style="margin: 0 0 25px 0; font-size: 1.2em; opacity: 0.95; line-height: 1.6;">
+                Experience the ultimate freedom in AI image creation! Generate and edit unlimited images without restrictions, 
+                including NSFW content, with our premium AI image editing platform.
+            </p>
+            
+            <div style="display: flex; justify-content: center; gap: 25px; flex-wrap: wrap; margin: 30px 0;">
+                <a href="https://omnicreator.net/#generator" target="_blank" style="
+                    display: inline-flex; 
+                    align-items: center;
+                    justify-content: center;
+                    padding: 20px 40px; 
+                    background: linear-gradient(135deg, #ff6b6b 0%, #feca57 100%);
+                    color: white; 
+                    text-decoration: none; 
+                    border-radius: 15px; 
+                    font-weight: 700;
+                    font-size: 18px;
+                    text-align: center;
+                    min-width: 250px;
+                    box-shadow: 0 8px 25px rgba(255, 107, 107, 0.4);
+                    transition: all 0.3s ease;
+                    border: none;
+                    transform: scale(1);
+                " onmouseover="this.style.transform='scale(1.05)'" onmouseout="this.style.transform='scale(1)'">
+                    &#128640; Get Unlimited Access Now
+                </a>
+                
+            </div>
+            
+            <p style="color: rgba(255,255,255,0.9); font-size: 1em; margin: 20px 0 0 0;">
+                Join thousands of creators who trust Omni Creator for unrestricted AI image generation!
+            </p>
+        </div>
+
+        <div style="text-align: center; margin: 25px auto; background: linear-gradient(135deg, #f5f7fa 0%, #c3cfe2 100%); padding: 35px; border-radius: 20px; box-shadow: 0 10px 30px rgba(0,0,0,0.1);">
+            <h2 style="color: #2c3e50; margin: 0 0 20px 0; font-size: 1.9em; font-weight: 700;">
+                &#11088; Professional AI Image Editor - No Restrictions
+            </h2>
+            <p style="color: #555; font-size: 1.1em; line-height: 1.6; margin: 0 0 20px 0; padding: 0 20px;">
+                Transform your creative vision into reality with our advanced AI image editing platform. Whether you're creating 
+                art, editing photos, designing content, or working with any type of imagery - our powerful AI removes all limitations 
+                and gives you complete creative freedom.
+            </p>
+        </div>
+
+        <div style="display: grid; grid-template-columns: repeat(auto-fit, minmax(300px, 1fr)); gap: 25px; margin: 40px 0;">
+            
+            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #e74c3c;">
+                <h3 style="color: #e74c3c; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
+                    &#127919; Unlimited Generation
+                </h3>
+                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
+                    Premium users enjoy unlimited image generation without daily limits, rate restrictions, or content barriers. 
+                    Create as many images as you need, whenever you need them.
+                </p>
+            </div>
+
+            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #3498db;">
+                <h3 style="color: #3498db; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
+                    🔓 No Content Restrictions
+                </h3>
+                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
+                    Generate and edit any type of content without NSFW filters or content limitations. Complete creative 
+                    freedom for artists, designers, and content creators.
+                </p>
+            </div>
+
+            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #27ae60;">
+                <h3 style="color: #27ae60; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
+                    &#9889; Lightning Fast Processing
+                </h3>
+                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
+                    Advanced AI infrastructure delivers high-quality results in seconds. No waiting in queues, 
+                    no processing delays - just instant, professional-grade image editing.
+                </p>
+            </div>
+
+            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #9b59b6;">
+                <h3 style="color: #9b59b6; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
+                    &#127912; Advanced Editing Tools
+                </h3>
+                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
+                    Global transformations, precision local editing, style transfer, object removal, background replacement, 
+                    and dozens of other professional editing capabilities.
+                </p>
+            </div>
+
+            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #f39c12;">
+                <h3 style="color: #f39c12; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
+                    &#128142; Premium Quality
+                </h3>
+                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
+                    State-of-the-art AI models trained on millions of images deliver exceptional quality and realism. 
+                    Professional results suitable for commercial use and high-end projects.
+                </p>
+            </div>
+
+            <div style="background: white; padding: 30px; border-radius: 15px; box-shadow: 0 5px 20px rgba(0,0,0,0.08); border-left: 5px solid #34495e;">
+                <h3 style="color: #34495e; margin: 0 0 15px 0; font-size: 1.4em; font-weight: 600;">
+                    🌍 Multi-Modal Support
+                </h3>
+                <p style="color: #666; margin: 0; line-height: 1.6; font-size: 1em;">
+                    Support for all image formats, styles, and use cases. From photorealistic portraits to artistic creations, 
+                    product photography to digital art - we handle everything.
+                </p>
+            </div>
+
+        </div>
+
+        <div style="background: linear-gradient(135deg, #11998e 0%, #38ef7d 100%); color: white; padding: 40px; border-radius: 20px; margin: 40px 0; text-align: center;">
+            <h2 style="margin: 0 0 25px 0; font-size: 1.8em; font-weight: 700;">
+                &#128142; Why Choose Omni Creator Premium?
+            </h2>
+            <div style="display: grid; grid-template-columns: repeat(auto-fit, minmax(250px, 1fr)); gap: 20px; margin: 30px 0;">
+                
+                <div style="background: rgba(255,255,255,0.15); padding: 20px; border-radius: 12px;">
+                    <h4 style="margin: 0 0 10px 0; font-size: 1.2em;">🚫 No Rate Limits</h4>
+                    <p style="margin: 0; opacity: 0.9; font-size: 0.95em;">Generate unlimited images without waiting periods or daily restrictions</p>
+                </div>
+
+                <div style="background: rgba(255,255,255,0.15); padding: 20px; border-radius: 12px;">
+                    <h4 style="margin: 0 0 10px 0; font-size: 1.2em;">🎭 Unrestricted Content</h4>
+                    <p style="margin: 0; opacity: 0.9; font-size: 0.95em;">Create any type of content without NSFW filters or censorship</p>
+                </div>
 
-    # Create and launch app
-    app = app_module.create_app()
+                <div style="background: rgba(255,255,255,0.15); padding: 20px; border-radius: 12px;">
+                    <h4 style="margin: 0 0 10px 0; font-size: 1.2em;">&#9889; Priority Processing</h4>
+                    <p style="margin: 0; opacity: 0.9; font-size: 0.95em;">Skip queues and get instant results with dedicated processing power</p>
+                </div>
+
+                <div style="background: rgba(255,255,255,0.15); padding: 20px; border-radius: 12px;">
+                    <h4 style="margin: 0 0 10px 0; font-size: 1.2em;">&#127912; Advanced Features</h4>
+                    <p style="margin: 0; opacity: 0.9; font-size: 0.95em;">Access to latest AI models and cutting-edge editing capabilities</p>
+                </div>
+
+            </div>
+            <div style="display: flex; justify-content: center; margin: 25px 0 0 0;">
+                <a href="https://omnicreator.net/#generator" target="_blank" style="
+                    display: inline-flex; 
+                    align-items: center;
+                    justify-content: center;
+                    padding: 18px 35px; 
+                    background: rgba(255,255,255,0.9);
+                    color: #333; 
+                    text-decoration: none; 
+                    border-radius: 15px; 
+                    font-weight: 700;
+                    font-size: 16px;
+                    text-align: center;
+                    min-width: 200px;
+                    box-shadow: 0 6px 20px rgba(0,0,0,0.3);
+                    transition: all 0.3s ease;
+                    border: none;
+                ">&#11088; Start Creating Now</a>
+            </div>
+        </div>
+
+        <div style="background: linear-gradient(135deg, #ff9a9e 0%, #fecfef 50%, #fecfef 100%); padding: 30px; border-radius: 15px; margin: 40px 0;">
+            <h3 style="color: #8b5cf6; text-align: center; margin: 0 0 25px 0; font-size: 1.5em; font-weight: 700;">
+                &#128161; Pro Tips for Best Results
+            </h3>
+            <div style="display: grid; grid-template-columns: repeat(auto-fit, minmax(300px, 1fr)); gap: 18px;">
+                
+                <div style="background: rgba(255,255,255,0.85); padding: 18px; border-radius: 12px;">
+                    <strong style="color: #8b5cf6; font-size: 1.1em;">📝 Clear Descriptions:</strong>
+                    <p style="color: #555; margin: 5px 0 0 0; line-height: 1.5;">Use detailed, specific prompts for better results. Describe colors, styles, lighting, and composition clearly.</p>
+                </div>
+
+                <div style="background: rgba(255,255,255,0.85); padding: 18px; border-radius: 12px;">
+                    <strong style="color: #8b5cf6; font-size: 1.1em;">&#127919; Local Editing:</strong>
+                    <p style="color: #555; margin: 5px 0 0 0; line-height: 1.5;">Use precise brush strokes to select areas for local editing. Smaller, focused edits often yield better results.</p>
+                </div>
+
+                <div style="background: rgba(255,255,255,0.85); padding: 18px; border-radius: 12px;">
+                    <strong style="color: #8b5cf6; font-size: 1.1em;">&#9889; Iterative Process:</strong>
+                    <p style="color: #555; margin: 5px 0 0 0; line-height: 1.5;">Use "Use as Input" feature to refine results. Multiple iterations can achieve complex transformations.</p>
+                </div>
+
+                <div style="background: rgba(255,255,255,0.85); padding: 18px; border-radius: 12px;">
+                    <strong style="color: #8b5cf6; font-size: 1.1em;">&#128444; Image Quality:</strong>
+                    <p style="color: #555; margin: 5px 0 0 0; line-height: 1.5;">Higher resolution input images (up to 10MB) generally produce better editing results and finer details.</p>
+                </div>
+
+            </div>
+        </div>
+
+        <div style="text-align: center; margin: 25px auto; background: linear-gradient(135deg, #a8edea 0%, #fed6e3 100%); padding: 35px; border-radius: 20px; box-shadow: 0 10px 30px rgba(0,0,0,0.1);">
+            <h2 style="color: #2c3e50; margin: 0 0 20px 0; font-size: 1.8em; font-weight: 700;">
+                &#128640; Perfect For Every Creative Need
+            </h2>
+            <div style="display: grid; grid-template-columns: repeat(auto-fit, minmax(200px, 1fr)); gap: 20px; margin: 25px 0; text-align: left;">
+                
+                <div style="background: rgba(255,255,255,0.8); padding: 20px; border-radius: 12px;">
+                    <h4 style="color: #e74c3c; margin: 0 0 10px 0;">🎨 Digital Art</h4>
+                    <ul style="color: #555; margin: 0; padding-left: 18px; line-height: 1.6;">
+                        <li>Character design</li>
+                        <li>Concept art</li>
+                        <li>Style transfer</li>
+                        <li>Artistic effects</li>
+                    </ul>
+                </div>
+
+                <div style="background: rgba(255,255,255,0.8); padding: 20px; border-radius: 12px;">
+                    <h4 style="color: #3498db; margin: 0 0 10px 0;">📸 Photography</h4>
+                    <ul style="color: #555; margin: 0; padding-left: 18px; line-height: 1.6;">
+                        <li>Background replacement</li>
+                        <li>Object removal</li>
+                        <li>Lighting adjustment</li>
+                        <li>Portrait enhancement</li>
+                    </ul>
+                </div>
+
+                <div style="background: rgba(255,255,255,0.8); padding: 20px; border-radius: 12px;">
+                    <h4 style="color: #27ae60; margin: 0 0 10px 0;">🛍️ E-commerce</h4>
+                    <ul style="color: #555; margin: 0; padding-left: 18px; line-height: 1.6;">
+                        <li>Product photography</li>
+                        <li>Lifestyle shots</li>
+                        <li>Color variations</li>
+                        <li>Context placement</li>
+                    </ul>
+                </div>
+
+                <div style="background: rgba(255,255,255,0.8); padding: 20px; border-radius: 12px;">
+                    <h4 style="color: #9b59b6; margin: 0 0 10px 0;">📱 Social Media</h4>
+                    <ul style="color: #555; margin: 0; padding-left: 18px; line-height: 1.6;">
+                        <li>Content creation</li>
+                        <li>Meme generation</li>
+                        <li>Brand visuals</li>
+                        <li>Viral content</li>
+                    </ul>
+                </div>
+
+            </div>
+            <div style="text-align: center; margin: 25px 0 0 0;">
+                <a href="https://omnicreator.net/#generator" target="_blank" style="
+                    display: inline-flex; 
+                    align-items: center;
+                    justify-content: center;
+                    padding: 18px 35px; 
+                    background: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
+                    color: white; 
+                    text-decoration: none; 
+                    border-radius: 15px; 
+                    font-weight: 700;
+                    font-size: 16px;
+                    text-align: center;
+                    min-width: 220px;
+                    box-shadow: 0 8px 25px rgba(102, 126, 234, 0.4);
+                    transition: all 0.3s ease;
+                    border: none;
+                ">🎯 Start Your Project Now</a>
+            </div>
+        </div>
+
+    </div>
+    
+    <div style="text-align: center; margin: 30px auto 20px auto; padding: 20px;">
+        <p style="margin: 0 0 10px 0; font-size: 18px; color: #333; font-weight: 500;">
+            Powered by <a href="https://omnicreator.net/#generator" target="_blank" style="color: #667eea; text-decoration: none; font-weight: bold;">Omni Creator</a>
+        </p>
+        <p style="margin: 0; font-size: 14px; color: #999; font-weight: 400;">
+            The ultimate AI image generation and editing platform • Unlimited creativity, zero restrictions
+        </p>
+    </div>
+    """, padding=False)
+    
+    return app
+
+if __name__ == "__main__":
+    app = create_app()
+    # Improve queue configuration to handle high concurrency and prevent SSE connection issues
     app.queue(
-        default_concurrency_limit=20,
-        max_size=50,
-        api_open=False
+        default_concurrency_limit=20,  # Default concurrency limit
+        max_size=50,  # Maximum queue size
+        api_open=False  # Close API access to reduce resource consumption
     )
     app.launch(
         server_name="0.0.0.0",
-        show_error=True,
-        quiet=False,
-        max_threads=40,
+        show_error=True,  # Show detailed error information
+        quiet=False,  # Keep log output
+        max_threads=40,  # Increase thread pool size
         height=800,
-        favicon_path=None
-    )
-
-except ImportError as e:
-    print(f"Failed to import compiled modules: {e}")
-    print("Make sure to run build_encrypted.py first to compile the modules")
-    import traceback
-    traceback.print_exc()
-    sys.exit(1)
-except Exception as e:
-    print(f"Error running app: {e}")
-    import traceback
-    traceback.print_exc()
-    sys.exit(1)
+        favicon_path=None  # Reduce resource loading
+    )

nfsw.py

@@ -0,0 +1,262 @@
+import os
+from PIL import Image
+import numpy as np
+import onnxruntime as ort
+import json
+from huggingface_hub import hf_hub_download
+
+
+class NSFWDetector:
+    """
+    NSFW检测器类，使用YOLOv9模型进行图像分类
+    """
+    
+    def __init__(self, repo_id="Falconsai/nsfw_image_detection", 
+                 model_filename="falconsai_yolov9_nsfw_model_quantized.pt",
+                 labels_filename="labels.json",
+                 input_size=(224, 224)):
+        """
+        初始化NSFW检测器
+        
+        Args:
+            repo_id (str): Hugging Face仓库ID
+            model_filename (str): 模型文件名
+            labels_filename (str): 标签文件名
+            input_size (tuple): 模型输入尺寸 (height, width)
+        """
+        self.repo_id = repo_id
+        self.model_filename = model_filename
+        self.labels_filename = labels_filename
+        self.input_size = input_size
+        
+        # 从Hugging Face下载文件
+        self.model_path = self._download_model()
+        self.labels_path = self._download_labels()
+        
+        # 加载标签
+        self.labels = self._load_labels()
+        
+        # 加载模型
+        self.session = self._load_model()
+        self.input_name = self.session.get_inputs()[0].name
+        self.output_name = self.session.get_outputs()[0].name
+        
+    def _download_model(self):
+        """
+        从Hugging Face下载模型文件
+        
+        Returns:
+            str: 下载的模型文件路径
+        """
+        try:
+            print(f"正在从 {self.repo_id} 下载模型文件: {self.model_filename}")
+            model_path = hf_hub_download(
+                repo_id=self.repo_id,
+                filename=self.model_filename,
+                cache_dir="./hf_cache"
+            )
+            print(f"✅ 模型下载成功: {model_path}")
+            return model_path
+        except Exception as e:
+            raise RuntimeError(f"模型下载失败: {e}")
+    
+    def _download_labels(self):
+        """
+        从Hugging Face下载标签文件
+        
+        Returns:
+            str: 下载的标签文件路径
+        """
+        try:
+            print(f"正在从 {self.repo_id} 下载标签文件: {self.labels_filename}")
+            labels_path = hf_hub_download(
+                repo_id=self.repo_id,
+                filename=self.labels_filename,
+                cache_dir="./hf_cache"
+            )
+            print(f"✅ 标签文件下载成功: {labels_path}")
+            return labels_path
+        except Exception as e:
+            raise RuntimeError(f"标签文件下载失败: {e}")
+    
+    def _load_labels(self):
+        """
+        加载类别标签
+        
+        Returns:
+            dict: 标签字典
+        """
+        try:
+            with open(self.labels_path, "r") as f:
+                return json.load(f)
+        except FileNotFoundError:
+            raise FileNotFoundError(f"标签文件未找到: {self.labels_path}")
+        except json.JSONDecodeError:
+            raise ValueError(f"标签文件格式错误: {self.labels_path}")
+    
+    def _load_model(self):
+        """
+        加载ONNX模型
+        
+        Returns:
+            onnxruntime.InferenceSession: 模型会话
+        """
+        try:
+            return ort.InferenceSession(self.model_path)
+        except Exception as e:
+            raise RuntimeError(f"模型加载失败: {self.model_path}, 错误: {e}")
+    
+    def _preprocess_image(self, image_path):
+        """
+        图像预处理
+        
+        Args:
+            image_path (str): 图像文件路径
+            
+        Returns:
+            tuple: (预处理后的张量, 原始图像)
+        """
+        try:
+            # 加载并转换图像
+            original_image = Image.open(image_path).convert("RGB")
+            
+            # 调整尺寸
+            image_resized = original_image.resize(self.input_size, Image.Resampling.BILINEAR)
+            
+            # 转换为numpy数组并归一化
+            image_np = np.array(image_resized, dtype=np.float32) / 255.0
+            
+            # 调整维度顺序 [H, W, C] -> [C, H, W]
+            image_np = np.transpose(image_np, (2, 0, 1))
+            
+            # 添加批次维度 [C, H, W] -> [1, C, H, W]
+            input_tensor = np.expand_dims(image_np, axis=0).astype(np.float32)
+            
+            return input_tensor, original_image
+            
+        except FileNotFoundError:
+            raise FileNotFoundError(f"图像文件未找到: {image_path}")
+        except Exception as e:
+            raise RuntimeError(f"图像预处理失败: {e}")
+    
+    def _postprocess_predictions(self, predictions):
+        """
+        后处理预测结果
+        
+        Args:
+            predictions: 模型预测输出
+            
+        Returns:
+            str: 预测的类别标签
+        """
+        predicted_index = np.argmax(predictions)
+        predicted_label = self.labels[str(predicted_index)]
+        return predicted_label
+    
+    def predict(self, image_path):
+        """
+        对单张图像进行NSFW检测
+        
+        Args:
+            image_path (str): 图像文件路径
+            
+        Returns:
+            tuple: (预测标签, 原始图像)
+        """
+        # 预处理图像
+        input_tensor, original_image = self._preprocess_image(image_path)
+        
+        # 运行推理
+        outputs = self.session.run([self.output_name], {self.input_name: input_tensor})
+        predictions = outputs[0]
+        
+        # 后处理结果
+        predicted_label = self._postprocess_predictions(predictions)
+        
+        return predicted_label, original_image
+    
+    def predict_label_only(self, image_path):
+        """
+        只返回预测标签（不返回图像）
+        
+        Args:
+            image_path (str): 图像文件路径
+            
+        Returns:
+            str: 预测的类别标签
+        """
+        predicted_label, _ = self.predict(image_path)
+        return predicted_label
+    
+    def predict_from_pil(self, pil_image):
+        """
+        直接从PIL Image对象进行NSFW检测
+        
+        Args:
+            pil_image (PIL.Image): PIL图像对象
+            
+        Returns:
+            tuple: (预测标签, 原始图像)
+        """
+        try:
+            # 确保是RGB格式
+            if pil_image.mode != "RGB":
+                pil_image = pil_image.convert("RGB")
+            
+            # 调整尺寸
+            image_resized = pil_image.resize(self.input_size, Image.Resampling.BILINEAR)
+            
+            # 转换为numpy数组并归一化
+            image_np = np.array(image_resized, dtype=np.float32) / 255.0
+            
+            # 调整维度顺序 [H, W, C] -> [C, H, W]
+            image_np = np.transpose(image_np, (2, 0, 1))
+            
+            # 添加批次维度 [C, H, W] -> [1, C, H, W]
+            input_tensor = np.expand_dims(image_np, axis=0).astype(np.float32)
+            
+            # 运行推理
+            outputs = self.session.run([self.output_name], {self.input_name: input_tensor})
+            predictions = outputs[0]
+            
+            # 后处理结果
+            predicted_label = self._postprocess_predictions(predictions)
+            
+            return predicted_label, pil_image
+            
+        except Exception as e:
+            raise RuntimeError(f"PIL图像预测失败: {e}")
+    
+    def predict_pil_label_only(self, pil_image):
+        """
+        从PIL Image对象只返回预测标签
+        
+        Args:
+            pil_image (PIL.Image): PIL图像对象
+            
+        Returns:
+            str: 预测的类别标签
+        """
+        predicted_label, _ = self.predict_from_pil(pil_image)
+        return predicted_label
+
+# --- 使用示例 ---
+if __name__ == "__main__":
+    # 配置参数
+    single_image_path = "datas/bad01.jpg"
+    
+    try:
+        # 创建检测器实例（自动从Hugging Face下载）
+        detector = NSFWDetector()
+        
+        # 检查图像文件是否存在
+        if os.path.exists(single_image_path):
+            # 进行预测
+            predicted_label = detector.predict_label_only(single_image_path)
+            print(f"图像文件: {single_image_path}")
+            print(f"预测结果: {predicted_label}")
+        else:
+            print(f"错误: 指定的图像文件不存在: {single_image_path}")
+            
+    except Exception as e:
+        print(f"初始化检测器时发生错误: {e}")

pipeline.py

@@ -1,1934 +0,0 @@
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from typing import Optional, Tuple, Union, List, Dict, Any, Callable
-from dataclasses import dataclass
-import numpy as np
-from PIL import Image
-import torchvision.transforms as T
-from torchvision.transforms.functional import to_tensor, normalize
-import warnings
-from contextlib import contextmanager
-from functools import wraps
-
-from transformers import PretrainedConfig, PreTrainedModel, CLIPTextModel, CLIPTokenizer
-from transformers.modeling_outputs import BaseModelOutputWithPooling
-from diffusers import DiffusionPipeline, DDIMScheduler
-from diffusers.configuration_utils import ConfigMixin, register_to_config
-from diffusers.models.modeling_utils import ModelMixin
-from diffusers.utils import BaseOutput
-
-# Optimization imports
-try:
-    import transformer_engine.pytorch as te
-    from transformer_engine.common import recipe
-    HAS_TRANSFORMER_ENGINE = True
-except ImportError:
-    HAS_TRANSFORMER_ENGINE = False
-
-try:
-    from torch._dynamo import config as dynamo_config
-    HAS_TORCH_COMPILE = hasattr(torch, 'compile')
-except ImportError:
-    HAS_TORCH_COMPILE = False
-
-# -----------------------------------------------------------------------------
-# 1. Advanced Configuration (8B Scale)
-# -----------------------------------------------------------------------------
-
-class OmniMMDitV2Config(PretrainedConfig):
-    model_type = "omnimm_dit_v2"
-    
-    def __init__(
-        self,
-        vocab_size: int = 49408,
-        hidden_size: int = 4096,          # 4096 dim for ~7B-8B scale
-        intermediate_size: int = 11008,   # Llama-style MLP expansion
-        num_hidden_layers: int = 32,      # Deep network
-        num_attention_heads: int = 32,
-        num_key_value_heads: Optional[int] = 8,  # GQA (Grouped Query Attention)
-        hidden_act: str = "silu",
-        max_position_embeddings: int = 4096,
-        initializer_range: float = 0.02,
-        rms_norm_eps: float = 1e-5,
-        use_cache: bool = True,
-        pad_token_id: int = 0,
-        bos_token_id: int = 1,
-        eos_token_id: int = 2,
-        tie_word_embeddings: bool = False,
-        rope_theta: float = 10000.0,
-        # DiT Specifics
-        patch_size: int = 2,
-        in_channels: int = 4,             # VAE Latent channels
-        out_channels: int = 4, # x2 for variance if learned
-        frequency_embedding_size: int = 256,
-        # Multi-Modal Specifics
-        max_condition_images: int = 3,    # Support 1-3 input images
-        visual_embed_dim: int = 1024,     # e.g., SigLIP or CLIP Vision
-        text_embed_dim: int = 4096,       # T5-XXL or similar
-        use_temporal_attention: bool = True, # For Video generation
-        # Optimization Configs
-        use_fp8_quantization: bool = False,
-        use_compilation: bool = False,
-        compile_mode: str = "reduce-overhead",
-        use_flash_attention: bool = True,
-        **kwargs,
-    ):
-        self.vocab_size = vocab_size
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.num_key_value_heads = num_key_value_heads
-        self.hidden_act = hidden_act
-        self.max_position_embeddings = max_position_embeddings
-        self.initializer_range = initializer_range
-        self.rms_norm_eps = rms_norm_eps
-        self.use_cache = use_cache
-        self.rope_theta = rope_theta
-        self.patch_size = patch_size
-        self.in_channels = in_channels
-        self.out_channels = out_channels
-        self.frequency_embedding_size = frequency_embedding_size
-        self.max_condition_images = max_condition_images
-        self.visual_embed_dim = visual_embed_dim
-        self.text_embed_dim = text_embed_dim
-        self.use_temporal_attention = use_temporal_attention
-        self.use_fp8_quantization = use_fp8_quantization
-        self.use_compilation = use_compilation
-        self.compile_mode = compile_mode
-        self.use_flash_attention = use_flash_attention
-        super().__init__(
-            pad_token_id=pad_token_id,
-            bos_token_id=bos_token_id,
-            eos_token_id=eos_token_id,
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs,
-        )
-
-# -----------------------------------------------------------------------------
-# 2. Professional Building Blocks (RoPE, SwiGLU, AdaLN)
-# -----------------------------------------------------------------------------
-
-class OmniRMSNorm(nn.Module):
-    def __init__(self, hidden_size, eps=1e-6):
-        super().__init__()
-        self.weight = nn.Parameter(torch.ones(hidden_size))
-        self.variance_epsilon = eps
-
-    def forward(self, hidden_states):
-        input_dtype = hidden_states.dtype
-        hidden_states = hidden_states.to(torch.float32)
-        variance = hidden_states.pow(2).mean(-1, keepdim=True)
-        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
-        return self.weight * hidden_states.to(input_dtype)
-
-class OmniRotaryEmbedding(nn.Module):
-    """Complex implementation of Rotary Positional Embeddings for DiT"""
-    def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
-        super().__init__()
-        self.dim = dim
-        self.max_position_embeddings = max_position_embeddings
-        self.base = base
-        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-
-    def forward(self, x, seq_len=None):
-        t = torch.arange(seq_len or x.shape[1], device=x.device).type_as(self.inv_freq)
-        freqs = torch.outer(t, self.inv_freq)
-        emb = torch.cat((freqs, freqs), dim=-1)
-        return emb.cos(), emb.sin()
-
-class OmniSwiGLU(nn.Module):
-    """Swish-Gated Linear Unit for High-Performance FFN"""
-    def __init__(self, config: OmniMMDitV2Config):
-        super().__init__()
-        self.w1 = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
-        self.w2 = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
-        self.w3 = nn.Linear(config.hidden_size, config.intermediate_size, bias=False)
-
-    def forward(self, x):
-        return self.w2(F.silu(self.w1(x)) * self.w3(x))
-
-class TimestepEmbedder(nn.Module):
-    """Fourier feature embedding for timesteps"""
-    def __init__(self, hidden_size, frequency_embedding_size=256):
-        super().__init__()
-        self.mlp = nn.Sequential(
-            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
-            nn.SiLU(),
-            nn.Linear(hidden_size, hidden_size, bias=True),
-        )
-        self.frequency_embedding_size = frequency_embedding_size
-
-    @staticmethod
-    def timestep_embedding(t, dim, max_period=10000):
-        half = dim // 2
-        freqs = torch.exp(
-            -torch.log(torch.tensor(max_period)) * torch.arange(start=0, end=half, dtype=torch.float32) / half
-        ).to(device=t.device)
-        args = t[:, None].float() * freqs[None]
-        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
-        if dim % 2:
-            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
-        return embedding
-
-    def forward(self, t, dtype):
-        t_freq = self.timestep_embedding(t, self.frequency_embedding_size).to(dtype)
-        return self.mlp(t_freq)
-
-# -----------------------------------------------------------------------------
-# 2.5. Data Processing Utilities
-# -----------------------------------------------------------------------------
-
-class OmniImageProcessor:
-    """Advanced image preprocessing for multi-modal diffusion models"""
-    
-    def __init__(
-        self,
-        image_mean: List[float] = [0.485, 0.456, 0.406],
-        image_std: List[float] = [0.229, 0.224, 0.225],
-        size: Tuple[int, int] = (512, 512),
-        interpolation: str = "bicubic",
-        do_normalize: bool = True,
-        do_center_crop: bool = False,
-    ):
-        self.image_mean = image_mean
-        self.image_std = image_std
-        self.size = size
-        self.do_normalize = do_normalize
-        self.do_center_crop = do_center_crop
-        
-        # Build transform pipeline
-        transforms_list = []
-        if do_center_crop:
-            transforms_list.append(T.CenterCrop(min(size)))
-        
-        interp_mode = {
-            "bilinear": T.InterpolationMode.BILINEAR,
-            "bicubic": T.InterpolationMode.BICUBIC,
-            "lanczos": T.InterpolationMode.LANCZOS,
-        }.get(interpolation, T.InterpolationMode.BICUBIC)
-        
-        transforms_list.append(T.Resize(size, interpolation=interp_mode, antialias=True))
-        self.transform = T.Compose(transforms_list)
-    
-    def preprocess(
-        self,
-        images: Union[Image.Image, np.ndarray, torch.Tensor, List[Union[Image.Image, np.ndarray, torch.Tensor]]],
-        return_tensors: str = "pt",
-    ) -> torch.Tensor:
-        """
-        Preprocess images for model input.
-        
-        Args:
-            images: Single image or list of images (PIL, numpy, or torch)
-            return_tensors: Return type ("pt" for PyTorch)
-            
-        Returns:
-            Preprocessed image tensor [B, C, H, W]
-        """
-        if not isinstance(images, list):
-            images = [images]
-        
-        processed = []
-        for img in images:
-            # Convert to PIL if needed
-            if isinstance(img, np.ndarray):
-                if img.dtype == np.uint8:
-                    img = Image.fromarray(img)
-                else:
-                    img = Image.fromarray((img * 255).astype(np.uint8))
-            elif isinstance(img, torch.Tensor):
-                img = T.ToPILImage()(img)
-            
-            # Apply transforms
-            img = self.transform(img)
-            
-            # Convert to tensor
-            if not isinstance(img, torch.Tensor):
-                img = to_tensor(img)
-            
-            # Normalize
-            if self.do_normalize:
-                img = normalize(img, self.image_mean, self.image_std)
-            
-            processed.append(img)
-        
-        # Stack into batch
-        if return_tensors == "pt":
-            return torch.stack(processed, dim=0)
-        
-        return processed
-    
-    def postprocess(
-        self,
-        images: torch.Tensor,
-        output_type: str = "pil",
-    ) -> Union[List[Image.Image], np.ndarray, torch.Tensor]:
-        """
-        Postprocess model output to desired format.
-        
-        Args:
-            images: Model output tensor [B, C, H, W]
-            output_type: "pil", "np", or "pt"
-            
-        Returns:
-            Processed images in requested format
-        """
-        # Denormalize if needed
-        if self.do_normalize:
-            mean = torch.tensor(self.image_mean).view(1, 3, 1, 1).to(images.device)
-            std = torch.tensor(self.image_std).view(1, 3, 1, 1).to(images.device)
-            images = images * std + mean
-        
-        # Clamp to valid range
-        images = torch.clamp(images, 0, 1)
-        
-        if output_type == "pil":
-            images = images.cpu().permute(0, 2, 3, 1).numpy()
-            images = (images * 255).round().astype(np.uint8)
-            return [Image.fromarray(img) for img in images]
-        elif output_type == "np":
-            return images.cpu().numpy()
-        else:
-            return images
-
-
-class OmniVideoProcessor:
-    """Video frame processing for temporal diffusion models"""
-    
-    def __init__(
-        self,
-        image_processor: OmniImageProcessor,
-        num_frames: int = 16,
-        frame_stride: int = 1,
-    ):
-        self.image_processor = image_processor
-        self.num_frames = num_frames
-        self.frame_stride = frame_stride
-    
-    def preprocess_video(
-        self,
-        video_frames: Union[List[Image.Image], np.ndarray, torch.Tensor],
-        temporal_interpolation: bool = True,
-    ) -> torch.Tensor:
-        """
-        Preprocess video frames for temporal model.
-        
-        Args:
-            video_frames: List of PIL images, numpy array [T, H, W, C], or tensor [T, C, H, W]
-            temporal_interpolation: Whether to interpolate to target frame count
-            
-        Returns:
-            Preprocessed video tensor [B, C, T, H, W]
-        """
-        # Convert to list of PIL images
-        if isinstance(video_frames, np.ndarray):
-            if video_frames.ndim == 4:  # [T, H, W, C]
-                video_frames = [Image.fromarray(frame) for frame in video_frames]
-            else:
-                raise ValueError(f"Expected 4D numpy array, got shape {video_frames.shape}")
-        elif isinstance(video_frames, torch.Tensor):
-            if video_frames.ndim == 4:  # [T, C, H, W]
-                video_frames = [T.ToPILImage()(frame) for frame in video_frames]
-            else:
-                raise ValueError(f"Expected 4D tensor, got shape {video_frames.shape}")
-        
-        # Sample frames if needed
-        total_frames = len(video_frames)
-        if temporal_interpolation and total_frames != self.num_frames:
-            indices = np.linspace(0, total_frames - 1, self.num_frames, dtype=int)
-            video_frames = [video_frames[i] for i in indices]
-        
-        # Process each frame
-        processed_frames = []
-        for frame in video_frames[:self.num_frames]:
-            frame_tensor = self.image_processor.preprocess(frame, return_tensors="pt")[0]
-            processed_frames.append(frame_tensor)
-        
-        # Stack: [T, C, H, W] -> [1, C, T, H, W]
-        video_tensor = torch.stack(processed_frames, dim=1).unsqueeze(0)
-        return video_tensor
-    
-    def postprocess_video(
-        self,
-        video_tensor: torch.Tensor,
-        output_type: str = "pil",
-    ) -> Union[List[Image.Image], np.ndarray, torch.Tensor]:
-        """
-        Postprocess video output.
-        
-        Args:
-            video_tensor: Model output [B, C, T, H, W] or [B, T, C, H, W]
-            output_type: "pil", "np", or "pt"
-            
-        Returns:
-            Processed video frames
-        """
-        # Normalize dimensions to [B, T, C, H, W]
-        if video_tensor.ndim == 5:
-            if video_tensor.shape[1] in [3, 4]:  # [B, C, T, H, W]
-                video_tensor = video_tensor.permute(0, 2, 1, 3, 4)
-        
-        batch_size, num_frames = video_tensor.shape[:2]
-        
-        # Process each frame
-        all_frames = []
-        for b in range(batch_size):
-            frames = []
-            for t in range(num_frames):
-                frame = video_tensor[b, t]  # [C, H, W]
-                frame = frame.unsqueeze(0)  # [1, C, H, W]
-                processed = self.image_processor.postprocess(frame, output_type=output_type)
-                frames.extend(processed)
-            all_frames.append(frames)
-        
-        return all_frames[0] if batch_size == 1 else all_frames
-
-
-class OmniLatentProcessor:
-    """VAE latent space encoding/decoding with scaling and normalization"""
-    
-    def __init__(
-        self,
-        vae: Any,
-        scaling_factor: float = 0.18215,
-        do_normalize_latents: bool = True,
-    ):
-        self.vae = vae
-        self.scaling_factor = scaling_factor
-        self.do_normalize_latents = do_normalize_latents
-    
-    @torch.no_grad()
-    def encode(
-        self,
-        images: torch.Tensor,
-        generator: Optional[torch.Generator] = None,
-        return_dict: bool = False,
-    ) -> torch.Tensor:
-        """
-        Encode images to latent space.
-        
-        Args:
-            images: Input images [B, C, H, W] in range [-1, 1]
-            generator: Random generator for sampling
-            return_dict: Whether to return dict or tensor
-            
-        Returns:
-            Latent codes [B, 4, H//8, W//8]
-        """
-        # VAE expects input in [-1, 1]
-        if images.min() >= 0:
-            images = images * 2.0 - 1.0
-        
-        # Encode
-        latent_dist = self.vae.encode(images).latent_dist
-        latents = latent_dist.sample(generator=generator)
-        
-        # Scale latents
-        latents = latents * self.scaling_factor
-        
-        # Additional normalization for stability
-        if self.do_normalize_latents:
-            latents = (latents - latents.mean()) / (latents.std() + 1e-6)
-        
-        return latents if not return_dict else {"latents": latents}
-    
-    @torch.no_grad()
-    def decode(
-        self,
-        latents: torch.Tensor,
-        return_dict: bool = False,
-    ) -> torch.Tensor:
-        """
-        Decode latents to image space.
-        
-        Args:
-            latents: Latent codes [B, 4, H//8, W//8]
-            return_dict: Whether to return dict or tensor
-            
-        Returns:
-            Decoded images [B, 3, H, W] in range [-1, 1]
-        """
-        # Denormalize if needed
-        if self.do_normalize_latents:
-            # Assume identity transform for simplicity in decoding
-            pass
-        
-        # Unscale
-        latents = latents / self.scaling_factor
-        
-        # Decode
-        images = self.vae.decode(latents).sample
-        
-        return images if not return_dict else {"images": images}
-    
-    @torch.no_grad()
-    def encode_video(
-        self,
-        video_frames: torch.Tensor,
-        generator: Optional[torch.Generator] = None,
-    ) -> torch.Tensor:
-        """
-        Encode video frames to latent space.
-        
-        Args:
-            video_frames: Input video [B, C, T, H, W] or [B, T, C, H, W]
-            generator: Random generator
-            
-        Returns:
-            Video latents [B, 4, T, H//8, W//8]
-        """
-        # Reshape to process frames independently
-        if video_frames.shape[2] not in [3, 4]:  # [B, T, C, H, W]
-            B, T, C, H, W = video_frames.shape
-            video_frames = video_frames.reshape(B * T, C, H, W)
-            
-            # Encode
-            latents = self.encode(video_frames, generator=generator)
-            
-            # Reshape back
-            latents = latents.reshape(B, T, *latents.shape[1:])
-            latents = latents.permute(0, 2, 1, 3, 4)  # [B, 4, T, H//8, W//8]
-        else:  # [B, C, T, H, W]
-            B, C, T, H, W = video_frames.shape
-            video_frames = video_frames.permute(0, 2, 1, 3, 4).reshape(B * T, C, H, W)
-            
-            latents = self.encode(video_frames, generator=generator)
-            latents = latents.reshape(B, T, *latents.shape[1:])
-            latents = latents.permute(0, 2, 1, 3, 4)
-        
-        return latents
-
-# -----------------------------------------------------------------------------
-# 3. Core Architecture: OmniMMDitBlock (3D-Attention + Modulation)
-# -----------------------------------------------------------------------------
-
-class OmniMMDitBlock(nn.Module):
-    def __init__(self, config: OmniMMDitV2Config, layer_idx: int):
-        super().__init__()
-        self.layer_idx = layer_idx
-        self.hidden_size = config.hidden_size
-        self.num_heads = config.num_attention_heads
-        self.head_dim = config.hidden_size // config.num_attention_heads
-        
-        # Self-Attention with QK-Norm
-        self.norm1 = OmniRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.attn = nn.MultiheadAttention(
-            config.hidden_size, config.num_attention_heads, batch_first=True
-        )
-        
-        self.q_norm = OmniRMSNorm(self.head_dim, eps=config.rms_norm_eps)
-        self.k_norm = OmniRMSNorm(self.head_dim, eps=config.rms_norm_eps)
-
-        # Cross-Attention for multimodal fusion
-        self.norm2 = OmniRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.cross_attn = nn.MultiheadAttention(
-            config.hidden_size, config.num_attention_heads, batch_first=True
-        )
-
-        # Feed-Forward Network with SwiGLU activation
-        self.norm3 = OmniRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-        self.ffn = OmniSwiGLU(config)
-
-        # Adaptive Layer Normalization with zero initialization
-        self.adaLN_modulation = nn.Sequential(
-            nn.SiLU(),
-            nn.Linear(config.hidden_size, 6 * config.hidden_size, bias=True)
-        )
-
-    def forward(
-        self, 
-        hidden_states: torch.Tensor, 
-        encoder_hidden_states: torch.Tensor, # Text embeddings
-        visual_context: Optional[torch.Tensor], # Reference image embeddings
-        timestep_emb: torch.Tensor,
-        rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
-    ) -> torch.Tensor:
-        
-        # AdaLN Modulation
-        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
-            self.adaLN_modulation(timestep_emb)[:, None].chunk(6, dim=-1)
-        )
-
-        # Self-Attention block
-        normed_hidden = self.norm1(hidden_states)
-        normed_hidden = normed_hidden * (1 + scale_msa) + shift_msa
-        
-        attn_output, _ = self.attn(normed_hidden, normed_hidden, normed_hidden)
-        hidden_states = hidden_states + gate_msa * attn_output
-
-        # Cross-Attention with multimodal conditioning
-        if visual_context is not None:
-             context = torch.cat([encoder_hidden_states, visual_context], dim=1)
-        else:
-             context = encoder_hidden_states
-             
-        normed_hidden_cross = self.norm2(hidden_states)
-        cross_output, _ = self.cross_attn(normed_hidden_cross, context, context)
-        hidden_states = hidden_states + cross_output
-
-        # Feed-Forward block
-        normed_ffn = self.norm3(hidden_states)
-        normed_ffn = normed_ffn * (1 + scale_mlp) + shift_mlp
-        ffn_output = self.ffn(normed_ffn)
-        hidden_states = hidden_states + gate_mlp * ffn_output
-
-        return hidden_states
-
-# -----------------------------------------------------------------------------
-# 4. The Model: OmniMMDitV2
-# -----------------------------------------------------------------------------
-
-class OmniMMDitV2(ModelMixin, PreTrainedModel):
-    """
-    Omni-Modal Multi-Dimensional Diffusion Transformer V2.
-    Supports: Text-to-Image, Image-to-Image (Edit), Image-to-Video.
-    """
-    config_class = OmniMMDitV2Config
-    _supports_gradient_checkpointing = True
-
-    def __init__(self, config: OmniMMDitV2Config):
-        super().__init__(config)
-        self.config = config
-
-        # Initialize optimizer for advanced features
-        self.optimizer = ModelOptimizer(
-            fp8_config=FP8Config(enabled=config.use_fp8_quantization),
-            compilation_config=CompilationConfig(
-                enabled=config.use_compilation,
-                mode=config.compile_mode,
-            ),
-            mixed_precision_config=MixedPrecisionConfig(
-                enabled=True,
-                dtype="bfloat16",
-            ),
-        )
-
-        # Input Latent Projection (Patchify)
-        self.x_embedder = nn.Linear(config.in_channels * config.patch_size * config.patch_size, config.hidden_size, bias=True)
-        
-        # Time & Vector Embeddings
-        self.t_embedder = TimestepEmbedder(config.hidden_size, config.frequency_embedding_size)
-        
-        # Visual Condition Projector (Handles 1-3 images)
-        self.visual_projector = nn.Sequential(
-            nn.Linear(config.visual_embed_dim, config.hidden_size),
-            nn.LayerNorm(config.hidden_size),
-            nn.Linear(config.hidden_size, config.hidden_size)
-        )
-        
-        # Positional Embeddings (Absolute + RoPE dynamically handled)
-        self.pos_embed = nn.Parameter(torch.zeros(1, config.max_position_embeddings, config.hidden_size), requires_grad=False)
-
-        # Transformer Backbone
-        self.blocks = nn.ModuleList([
-            OmniMMDitBlock(config, i) for i in range(config.num_hidden_layers)
-        ])
-        
-        # Final Layer (AdaLN-Zero + Linear)
-        self.final_layer = nn.Sequential(
-            OmniRMSNorm(config.hidden_size, eps=config.rms_norm_eps),
-            nn.Linear(config.hidden_size, config.patch_size * config.patch_size * config.out_channels, bias=True)
-        )
-        
-        self.initialize_weights()
-        
-        # Apply optimizations if enabled
-        if config.use_fp8_quantization or config.use_compilation:
-            self._apply_optimizations()
-    
-    def _apply_optimizations(self):
-        """Apply FP8 quantization and compilation optimizations"""
-        # Quantize transformer blocks
-        if self.config.use_fp8_quantization:
-            for i, block in enumerate(self.blocks):
-                self.blocks[i] = self.optimizer.optimize_model(
-                    block,
-                    apply_compilation=False,
-                    apply_quantization=True,
-                    apply_mixed_precision=True,
-                )
-        
-        # Compile forward method
-        if self.config.use_compilation and HAS_TORCH_COMPILE:
-            self.forward = torch.compile(
-                self.forward,
-                mode=self.config.compile_mode,
-                dynamic=True,
-            )
-
-    def initialize_weights(self):
-        def _basic_init(module):
-            if isinstance(module, nn.Linear):
-                torch.nn.init.xavier_uniform_(module.weight)
-                if module.bias is not None:
-                    nn.init.constant_(module.bias, 0)
-        self.apply(_basic_init)
-
-    def unpatchify(self, x, h, w):
-        c = self.config.out_channels
-        p = self.config.patch_size
-        h_ = h // p
-        w_ = w // p
-        x = x.reshape(shape=(x.shape[0], h_, w_, p, p, c))
-        x = torch.einsum('nhwpqc->nchpwq', x)
-        imgs = x.reshape(shape=(x.shape[0], c, h, w))
-        return imgs
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor, # Noisy Latents [B, C, H, W] or [B, C, F, H, W]
-        timestep: torch.LongTensor,
-        encoder_hidden_states: torch.Tensor, # Text Embeddings
-        visual_conditions: Optional[List[torch.Tensor]] = None, # List of [B, L, D]
-        video_frames: Optional[int] = None, # If generating video
-        return_dict: bool = True,
-    ) -> Union[torch.Tensor, BaseOutput]:
-        
-        batch_size, channels, _, _ = hidden_states.shape
-        
-        # Patchify input latents
-        p = self.config.patch_size
-        h, w = hidden_states.shape[-2], hidden_states.shape[-1]
-        x = hidden_states.unfold(2, p, p).unfold(3, p, p)
-        x = x.permute(0, 2, 3, 1, 4, 5).contiguous()
-        x = x.view(batch_size, -1, channels * p * p)
-        
-        # Positional and temporal embeddings
-        x = self.x_embedder(x)
-        x = x + self.pos_embed[:, :x.shape[1], :]
-        
-        t = self.t_embedder(timestep, x.dtype)
-        
-        # Process visual conditioning
-        visual_emb = None
-        if visual_conditions is not None:
-            concat_visuals = torch.cat(visual_conditions, dim=1)
-            visual_emb = self.visual_projector(concat_visuals)
-
-        # Transformer blocks
-        for block in self.blocks:
-            x = block(
-                hidden_states=x, 
-                encoder_hidden_states=encoder_hidden_states,
-                visual_context=visual_emb,
-                timestep_emb=t
-            )
-
-        # Output projection
-        x = self.final_layer[0](x)
-        x = self.final_layer[1](x)
-        
-        # Unpatchify to image space
-        output = self.unpatchify(x, h, w)
-
-        if not return_dict:
-            return (output,)
-
-        return BaseOutput(sample=output)
-
-# -----------------------------------------------------------------------------
-# 5. The "Fancy" Pipeline
-# -----------------------------------------------------------------------------
-
-class OmniMMDitV2Pipeline(DiffusionPipeline):
-    """
-    Omni-Modal Diffusion Transformer Pipeline.
-    
-    Supports text-guided image editing and video generation with
-    multi-image conditioning and advanced guidance techniques.
-    """
-    model: OmniMMDitV2
-    tokenizer: CLIPTokenizer
-    text_encoder: CLIPTextModel
-    vae: Any # AutoencoderKL
-    scheduler: DDIMScheduler
-    
-    _optional_components = ["visual_encoder"]
-
-    def __init__(
-        self,
-        model: OmniMMDitV2,
-        vae: Any,
-        text_encoder: CLIPTextModel,
-        tokenizer: CLIPTokenizer,
-        scheduler: DDIMScheduler,
-        visual_encoder: Optional[Any] = None,
-    ):
-        super().__init__()
-        self.register_modules(
-            model=model,
-            vae=vae,
-            text_encoder=text_encoder,
-            tokenizer=tokenizer,
-            scheduler=scheduler,
-            visual_encoder=visual_encoder
-        )
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
-        
-        # Initialize data processors
-        self.image_processor = OmniImageProcessor(
-            size=(512, 512),
-            interpolation="bicubic",
-            do_normalize=True,
-        )
-        self.video_processor = OmniVideoProcessor(
-            image_processor=self.image_processor,
-            num_frames=16,
-        )
-        self.latent_processor = OmniLatentProcessor(
-            vae=vae,
-            scaling_factor=0.18215,
-        )
-        
-        # Initialize model optimizer
-        self.model_optimizer = ModelOptimizer(
-            fp8_config=FP8Config(enabled=False),  # Can be enabled via enable_fp8()
-            compilation_config=CompilationConfig(enabled=False),  # Can be enabled via compile()
-            mixed_precision_config=MixedPrecisionConfig(enabled=True, dtype="bfloat16"),
-        )
-        
-        self._is_compiled = False
-        self._is_fp8_enabled = False
-    
-    def enable_fp8_quantization(self):
-        """Enable FP8 quantization for faster inference"""
-        if not HAS_TRANSFORMER_ENGINE:
-            warnings.warn("Transformer Engine not available. Install with: pip install transformer-engine")
-            return self
-        
-        self.model_optimizer.fp8_config.enabled = True
-        self.model = self.model_optimizer.optimize_model(
-            self.model,
-            apply_compilation=False,
-            apply_quantization=True,
-            apply_mixed_precision=False,
-        )
-        self._is_fp8_enabled = True
-        return self
-    
-    def compile_model(
-        self,
-        mode: str = "reduce-overhead",
-        fullgraph: bool = False,
-        dynamic: bool = True,
-    ):
-        """
-        Compile model using torch.compile for faster inference.
-        
-        Args:
-            mode: Compilation mode - "default", "reduce-overhead", "max-autotune"
-            fullgraph: Whether to compile the entire model as one graph
-            dynamic: Whether to enable dynamic shapes
-        """
-        if not HAS_TORCH_COMPILE:
-            warnings.warn("torch.compile not available. Upgrade to PyTorch 2.0+")
-            return self
-        
-        self.model_optimizer.compilation_config = CompilationConfig(
-            enabled=True,
-            mode=mode,
-            fullgraph=fullgraph,
-            dynamic=dynamic,
-        )
-        
-        self.model = self.model_optimizer._compile_model(self.model)
-        self._is_compiled = True
-        return self
-    
-    def enable_optimizations(
-        self,
-        enable_fp8: bool = False,
-        enable_compilation: bool = False,
-        compilation_mode: str = "reduce-overhead",
-    ):
-        """
-        Enable all optimizations at once.
-        
-        Args:
-            enable_fp8: Enable FP8 quantization
-            enable_compilation: Enable torch.compile
-            compilation_mode: Compilation mode for torch.compile
-        """
-        if enable_fp8:
-            self.enable_fp8_quantization()
-        
-        if enable_compilation:
-            self.compile_model(mode=compilation_mode)
-        
-        return self
-
-    @torch.no_grad()
-    def __call__(
-        self,
-        prompt: Union[str, List[str]] = None,
-        input_images: Optional[List[Union[torch.Tensor, Any]]] = None,
-        height: Optional[int] = 1024,
-        width: Optional[int] = 1024,
-        num_frames: Optional[int] = 1,
-        num_inference_steps: int = 50,
-        guidance_scale: float = 7.5,
-        image_guidance_scale: float = 1.5,
-        negative_prompt: Optional[Union[str, List[str]]] = None,
-        eta: float = 0.0,
-        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
-        latents: Optional[torch.Tensor] = None,
-        output_type: Optional[str] = "pil",
-        return_dict: bool = True,
-        callback: Optional[Callable[[int, int, torch.Tensor], None]] = None,
-        callback_steps: int = 1,
-        use_optimized_inference: bool = True,
-        **kwargs,
-    ):
-        # Use optimized inference context
-        with optimized_inference_mode(
-            enable_cudnn_benchmark=use_optimized_inference,
-            enable_tf32=use_optimized_inference,
-            enable_flash_sdp=use_optimized_inference,
-        ):
-            return self._forward_impl(
-                prompt=prompt,
-                input_images=input_images,
-                height=height,
-                width=width,
-                num_frames=num_frames,
-                num_inference_steps=num_inference_steps,
-                guidance_scale=guidance_scale,
-                image_guidance_scale=image_guidance_scale,
-                negative_prompt=negative_prompt,
-                eta=eta,
-                generator=generator,
-                latents=latents,
-                output_type=output_type,
-                return_dict=return_dict,
-                callback=callback,
-                callback_steps=callback_steps,
-                **kwargs,
-            )
-    
-    def _forward_impl(
-        self,
-        prompt: Union[str, List[str]] = None,
-        input_images: Optional[List[Union[torch.Tensor, Any]]] = None,
-        height: Optional[int] = 1024,
-        width: Optional[int] = 1024,
-        num_frames: Optional[int] = 1,
-        num_inference_steps: int = 50,
-        guidance_scale: float = 7.5,
-        image_guidance_scale: float = 1.5,
-        negative_prompt: Optional[Union[str, List[str]]] = None,
-        eta: float = 0.0,
-        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
-        latents: Optional[torch.Tensor] = None,
-        output_type: Optional[str] = "pil",
-        return_dict: bool = True,
-        callback: Optional[Callable[[int, int, torch.Tensor], None]] = None,
-        callback_steps: int = 1,
-        **kwargs,
-    ):
-        # Validate and set default dimensions
-        height = height or self.model.config.sample_size * self.vae_scale_factor
-        width = width or self.model.config.sample_size * self.vae_scale_factor
-
-        # Encode text prompts
-        if isinstance(prompt, str):
-            prompt = [prompt]
-        batch_size = len(prompt)
-        
-        text_inputs = self.tokenizer(
-            prompt, padding="max_length", max_length=self.tokenizer.model_max_length, truncation=True, return_tensors="pt"
-        )
-        text_embeddings = self.text_encoder(text_inputs.input_ids.to(self.device))[0]
-
-        # Encode visual conditions with preprocessing
-        visual_embeddings_list = []
-        if input_images:
-            if not isinstance(input_images, list):
-                input_images = [input_images]
-            if len(input_images) > 3:
-                raise ValueError("Maximum 3 reference images supported")
-            
-            for img in input_images:
-                # Preprocess image
-                if not isinstance(img, torch.Tensor):
-                    img_tensor = self.image_processor.preprocess(img, return_tensors="pt")
-                else:
-                    img_tensor = img
-                
-                img_tensor = img_tensor.to(device=self.device, dtype=text_embeddings.dtype)
-                
-                # Encode with visual encoder
-                if self.visual_encoder is not None:
-                    vis_emb = self.visual_encoder(img_tensor).last_hidden_state
-                else:
-                    # Fallback: use VAE encoder + projection
-                    with torch.no_grad():
-                        latent_features = self.vae.encode(img_tensor * 2 - 1).latent_dist.mode()
-                        B, C, H, W = latent_features.shape
-                        # Flatten spatial dims and project
-                        vis_emb = latent_features.flatten(2).transpose(1, 2)  # [B, H*W, C]
-                        # Simple projection to visual_embed_dim
-                        if vis_emb.shape[-1] != self.model.config.visual_embed_dim:
-                            proj = nn.Linear(vis_emb.shape[-1], self.model.config.visual_embed_dim).to(self.device)
-                            vis_emb = proj(vis_emb)
-                
-                visual_embeddings_list.append(vis_emb)
-
-        # Prepare timesteps
-        self.scheduler.set_timesteps(num_inference_steps, device=self.device)
-        timesteps = self.scheduler.timesteps
-
-        # Initialize latent space
-        num_channels_latents = self.model.config.in_channels
-        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
-        if num_frames > 1:
-            shape = (batch_size, num_channels_latents, num_frames, height // self.vae_scale_factor, width // self.vae_scale_factor)
-
-        latents = torch.randn(shape, generator=generator, device=self.device, dtype=text_embeddings.dtype)
-        latents = latents * self.scheduler.init_noise_sigma
-
-        # Denoising loop with optimizations
-        with self.progress_bar(total=num_inference_steps) as progress_bar:
-            for i, t in enumerate(timesteps):
-                latent_model_input = torch.cat([latents] * 2) if guidance_scale > 1.0 else latents
-                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
-                
-                # Use mixed precision autocast
-                with self.model_optimizer.autocast_context():
-                    noise_pred = self.model(
-                        hidden_states=latent_model_input,
-                        timestep=t,
-                        encoder_hidden_states=torch.cat([text_embeddings] * 2),
-                        visual_conditions=visual_embeddings_list * 2 if visual_embeddings_list else None,
-                        video_frames=num_frames
-                    ).sample
-
-                # Apply classifier-free guidance
-                if guidance_scale > 1.0:
-                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
-                
-                latents = self.scheduler.step(noise_pred, t, latents, eta=eta).prev_sample
-                
-                # Call callback if provided
-                if callback is not None and i % callback_steps == 0:
-                    callback(i, t, latents)
-                
-                progress_bar.update()
-
-        # Decode latents with proper post-processing
-        if output_type == "latent":
-            output_images = latents
-        else:
-            # Decode latents to pixel space
-            with torch.no_grad():
-                if num_frames > 1:
-                    # Video decoding: process frame by frame
-                    B, C, T, H, W = latents.shape
-                    latents_2d = latents.permute(0, 2, 1, 3, 4).reshape(B * T, C, H, W)
-                    decoded = self.latent_processor.decode(latents_2d)
-                    decoded = decoded.reshape(B, T, 3, H * 8, W * 8)
-                    
-                    # Convert to [0, 1] range
-                    decoded = (decoded / 2 + 0.5).clamp(0, 1)
-                    
-                    # Post-process video
-                    if output_type == "pil":
-                        output_images = self.video_processor.postprocess_video(decoded, output_type="pil")
-                    elif output_type == "np":
-                        output_images = decoded.cpu().numpy()
-                    else:
-                        output_images = decoded
-                else:
-                    # Image decoding
-                    decoded = self.latent_processor.decode(latents)
-                    decoded = (decoded / 2 + 0.5).clamp(0, 1)
-                    
-                    # Post-process images
-                    if output_type == "pil":
-                        output_images = self.image_processor.postprocess(decoded, output_type="pil")
-                    elif output_type == "np":
-                        output_images = decoded.cpu().numpy()
-                    else:
-                        output_images = decoded
-
-        if not return_dict:
-            return (output_images,)
-
-        return BaseOutput(images=output_images)
-
-# -----------------------------------------------------------------------------
-# 6. Advanced Multi-Modal Window Attention Block (Audio + Video + Image)
-# -----------------------------------------------------------------------------
-
-@dataclass
-class MultiModalInput:
-    """Container for multi-modal inputs"""
-    image_embeds: Optional[torch.Tensor] = None      # [B, L_img, D]
-    video_embeds: Optional[torch.Tensor] = None      # [B, T_video, L_vid, D]
-    audio_embeds: Optional[torch.Tensor] = None      # [B, T_audio, L_aud, D]
-    attention_mask: Optional[torch.Tensor] = None    # [B, total_length]
-
-
-class TemporalWindowPartition(nn.Module):
-    """
-    Partition temporal sequences into windows for efficient attention.
-    Supports both uniform and adaptive windowing strategies.
-    """
-    def __init__(
-        self,
-        window_size: int = 8,
-        shift_size: int = 0,
-        use_adaptive_window: bool = False,
-    ):
-        super().__init__()
-        self.window_size = window_size
-        self.shift_size = shift_size
-        self.use_adaptive_window = use_adaptive_window
-    
-    def partition(self, x: torch.Tensor) -> Tuple[torch.Tensor, Dict[str, Any]]:
-        """
-        Partition sequence into windows.
-        
-        Args:
-            x: Input tensor [B, T, L, D] or [B, L, D]
-            
-        Returns:
-            windowed: [B * num_windows, window_size, L, D]
-            info: Dictionary with partition information
-        """
-        if x.ndim == 3:  # Static input (image)
-            return x, {"is_temporal": False, "original_shape": x.shape}
-        
-        B, T, L, D = x.shape
-        
-        # Apply temporal shift for shifted window attention (Swin-Transformer style)
-        if self.shift_size > 0:
-            x = torch.roll(x, shifts=-self.shift_size, dims=1)
-        
-        # Pad if necessary
-        pad_t = (self.window_size - T % self.window_size) % self.window_size
-        if pad_t > 0:
-            x = F.pad(x, (0, 0, 0, 0, 0, pad_t))
-        
-        T_padded = T + pad_t
-        num_windows = T_padded // self.window_size
-        
-        # Reshape into windows: [B, num_windows, window_size, L, D]
-        x_windowed = x.view(B, num_windows, self.window_size, L, D)
-        
-        # Merge batch and window dims: [B * num_windows, window_size, L, D]
-        x_windowed = x_windowed.view(B * num_windows, self.window_size, L, D)
-        
-        info = {
-            "is_temporal": True,
-            "original_shape": (B, T, L, D),
-            "num_windows": num_windows,
-            "pad_t": pad_t,
-        }
-        
-        return x_windowed, info
-    
-    def merge(self, x_windowed: torch.Tensor, info: Dict[str, Any]) -> torch.Tensor:
-        """
-        Merge windows back to original sequence.
-        
-        Args:
-            x_windowed: Windowed tensor [B * num_windows, window_size, L, D]
-            info: Partition information from partition()
-            
-        Returns:
-            x: Merged tensor [B, T, L, D] or [B, L, D]
-        """
-        if not info["is_temporal"]:
-            return x_windowed
-        
-        B, T, L, D = info["original_shape"]
-        num_windows = info["num_windows"]
-        pad_t = info["pad_t"]
-        
-        # Reshape: [B * num_windows, window_size, L, D] -> [B, num_windows, window_size, L, D]
-        x = x_windowed.view(B, num_windows, self.window_size, L, D)
-        
-        # Merge windows: [B, T_padded, L, D]
-        x = x.view(B, num_windows * self.window_size, L, D)
-        
-        # Remove padding
-        if pad_t > 0:
-            x = x[:, :-pad_t, :, :]
-        
-        # Reverse temporal shift
-        if self.shift_size > 0:
-            x = torch.roll(x, shifts=self.shift_size, dims=1)
-        
-        return x
-
-
-class WindowCrossAttention(nn.Module):
-    """
-    Window-based Cross Attention with support for temporal sequences.
-    Performs attention within local windows for computational efficiency.
-    """
-    def __init__(
-        self,
-        dim: int,
-        num_heads: int = 8,
-        window_size: int = 8,
-        qkv_bias: bool = True,
-        attn_drop: float = 0.0,
-        proj_drop: float = 0.0,
-        use_relative_position_bias: bool = True,
-    ):
-        super().__init__()
-        self.dim = dim
-        self.num_heads = num_heads
-        self.window_size = window_size
-        self.head_dim = dim // num_heads
-        self.scale = self.head_dim ** -0.5
-        
-        # Query, Key, Value projections
-        self.q_proj = nn.Linear(dim, dim, bias=qkv_bias)
-        self.k_proj = nn.Linear(dim, dim, bias=qkv_bias)
-        self.v_proj = nn.Linear(dim, dim, bias=qkv_bias)
-        
-        # QK Normalization for stability
-        self.q_norm = OmniRMSNorm(self.head_dim)
-        self.k_norm = OmniRMSNorm(self.head_dim)
-        
-        # Attention dropout
-        self.attn_drop = nn.Dropout(attn_drop)
-        
-        # Output projection
-        self.proj = nn.Linear(dim, dim)
-        self.proj_drop = nn.Dropout(proj_drop)
-        
-        # Relative position bias (for temporal coherence)
-        self.use_relative_position_bias = use_relative_position_bias
-        if use_relative_position_bias:
-            # Temporal relative position bias
-            self.relative_position_bias_table = nn.Parameter(
-                torch.zeros((2 * window_size - 1), num_heads)
-            )
-            nn.init.trunc_normal_(self.relative_position_bias_table, std=0.02)
-            
-            # Get relative position index
-            coords = torch.arange(window_size)
-            relative_coords = coords[:, None] - coords[None, :]  # [window_size, window_size]
-            relative_coords += window_size - 1  # Shift to start from 0
-            self.register_buffer("relative_position_index", relative_coords)
-    
-    def get_relative_position_bias(self, window_size: int) -> torch.Tensor:
-        """Generate relative position bias for attention"""
-        if not self.use_relative_position_bias:
-            return None
-        
-        relative_position_bias = self.relative_position_bias_table[
-            self.relative_position_index[:window_size, :window_size].reshape(-1)
-        ].reshape(window_size, window_size, -1)
-        
-        # Permute to [num_heads, window_size, window_size]
-        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()
-        return relative_position_bias
-    
-    def forward(
-        self,
-        query: torch.Tensor,      # [B, T_q, L_q, D] or [B, L_q, D]
-        key: torch.Tensor,        # [B, T_k, L_k, D] or [B, L_k, D]
-        value: torch.Tensor,      # [B, T_v, L_v, D] or [B, L_v, D]
-        attention_mask: Optional[torch.Tensor] = None,
-    ) -> torch.Tensor:
-        """
-        Perform windowed cross attention.
-        
-        Args:
-            query: Query tensor
-            key: Key tensor  
-            value: Value tensor
-            attention_mask: Optional attention mask
-            
-        Returns:
-            Output tensor with same shape as query
-        """
-        # Handle both temporal and non-temporal inputs
-        is_temporal = query.ndim == 4
-        
-        if is_temporal:
-            B, T_q, L_q, D = query.shape
-            _, T_k, L_k, _ = key.shape
-            
-            # Flatten temporal and spatial dims for cross attention
-            query_flat = query.reshape(B, T_q * L_q, D)
-            key_flat = key.reshape(B, T_k * L_k, D)
-            value_flat = value.reshape(B, T_k * L_k, D)
-        else:
-            B, L_q, D = query.shape
-            _, L_k, _ = key.shape
-            query_flat = query
-            key_flat = key
-            value_flat = value
-        
-        # Project to Q, K, V
-        q = self.q_proj(query_flat)  # [B, N_q, D]
-        k = self.k_proj(key_flat)    # [B, N_k, D]
-        v = self.v_proj(value_flat)  # [B, N_v, D]
-        
-        # Reshape for multi-head attention
-        q = q.reshape(B, -1, self.num_heads, self.head_dim).transpose(1, 2)  # [B, H, N_q, head_dim]
-        k = k.reshape(B, -1, self.num_heads, self.head_dim).transpose(1, 2)  # [B, H, N_k, head_dim]
-        v = v.reshape(B, -1, self.num_heads, self.head_dim).transpose(1, 2)  # [B, H, N_v, head_dim]
-        
-        # Apply QK normalization
-        q = self.q_norm(q)
-        k = self.k_norm(k)
-        
-        # Scaled dot-product attention
-        attn = (q @ k.transpose(-2, -1)) * self.scale  # [B, H, N_q, N_k]
-        
-        # Add relative position bias if temporal
-        if is_temporal and self.use_relative_position_bias:
-            # Apply per-window bias
-            rel_bias = self.get_relative_position_bias(min(T_q, self.window_size))
-            if rel_bias is not None:
-                # Broadcast bias across spatial dimensions
-                attn = attn + rel_bias.unsqueeze(0).unsqueeze(2)
-        
-        # Apply attention mask
-        if attention_mask is not None:
-            attn = attn.masked_fill(attention_mask.unsqueeze(1).unsqueeze(2) == 0, float('-inf'))
-        
-        # Softmax and dropout
-        attn = F.softmax(attn, dim=-1)
-        attn = self.attn_drop(attn)
-        
-        # Apply attention to values
-        out = (attn @ v).transpose(1, 2).reshape(B, -1, D)  # [B, N_q, D]
-        
-        # Output projection
-        out = self.proj(out)
-        out = self.proj_drop(out)
-        
-        # Reshape back to original shape
-        if is_temporal:
-            out = out.reshape(B, T_q, L_q, D)
-        else:
-            out = out.reshape(B, L_q, D)
-        
-        return out
-
-
-class MultiModalFusionLayer(nn.Module):
-    """
-    Fuses multiple modalities (audio, video, image) with learnable fusion weights.
-    """
-    def __init__(
-        self,
-        dim: int,
-        num_modalities: int = 3,
-        fusion_type: str = "weighted",  # "weighted", "gated", "adaptive"
-    ):
-        super().__init__()
-        self.dim = dim
-        self.num_modalities = num_modalities
-        self.fusion_type = fusion_type
-        
-        if fusion_type == "weighted":
-            # Learnable fusion weights
-            self.fusion_weights = nn.Parameter(torch.ones(num_modalities) / num_modalities)
-        
-        elif fusion_type == "gated":
-            # Gated fusion with cross-modal interactions
-            self.gate_proj = nn.Sequential(
-                nn.Linear(dim * num_modalities, dim * 2),
-                nn.GELU(),
-                nn.Linear(dim * 2, num_modalities),
-                nn.Softmax(dim=-1)
-            )
-        
-        elif fusion_type == "adaptive":
-            # Adaptive fusion with per-token gating
-            self.adaptive_gate = nn.Sequential(
-                nn.Linear(dim, dim // 2),
-                nn.GELU(),
-                nn.Linear(dim // 2, num_modalities),
-                nn.Sigmoid()
-            )
-    
-    def forward(self, modality_features: List[torch.Tensor]) -> torch.Tensor:
-        """
-        Fuse multiple modality features.
-        
-        Args:
-            modality_features: List of [B, L, D] tensors for each modality
-            
-        Returns:
-            fused: Fused features [B, L, D]
-        """
-        if self.fusion_type == "weighted":
-            # Simple weighted sum
-            weights = F.softmax(self.fusion_weights, dim=0)
-            fused = sum(w * feat for w, feat in zip(weights, modality_features))
-        
-        elif self.fusion_type == "gated":
-            # Concatenate and compute gates
-            concat_features = torch.cat(modality_features, dim=-1)  # [B, L, D * num_modalities]
-            gates = self.gate_proj(concat_features)  # [B, L, num_modalities]
-            
-            # Apply gates
-            stacked = torch.stack(modality_features, dim=-1)  # [B, L, D, num_modalities]
-            fused = (stacked * gates.unsqueeze(2)).sum(dim=-1)  # [B, L, D]
-        
-        elif self.fusion_type == "adaptive":
-            # Adaptive per-token fusion
-            fused_list = []
-            for feat in modality_features:
-                gate = self.adaptive_gate(feat)  # [B, L, num_modalities]
-                fused_list.append(feat.unsqueeze(-1) * gate.unsqueeze(2))
-            
-            fused = torch.cat(fused_list, dim=-1).sum(dim=-1)  # [B, L, D]
-        
-        return fused
-
-
-class FancyMultiModalWindowAttentionBlock(nn.Module):
-    """
-    🎯 Fancy Multi-Modal Window Attention Block
-    
-    A state-of-the-art block that processes audio, video, and image embeddings
-    with temporal window-based cross-attention for efficient multi-modal fusion.
-    
-    Features:
-    - ✨ Temporal windowing for audio and video (frame-by-frame processing)
-    - 🪟 Shifted window attention for better temporal coherence (Swin-style)
-    - 🔄 Cross-modal attention between all modality pairs
-    - 🎭 Adaptive multi-modal fusion with learnable gates
-    - 🚀 Efficient computation with window partitioning
-    - 💎 QK normalization for training stability
-    
-    Architecture:
-        1. Temporal Partitioning (audio/video frames → windows)
-        2. Intra-Modal Self-Attention (within each modality)
-        3. Inter-Modal Cross-Attention (audio ↔ video ↔ image)
-        4. Multi-Modal Fusion (adaptive weighted combination)
-        5. Feed-Forward Network (SwiGLU activation)
-        6. Window Merging (reconstruct temporal sequences)
-    """
-    
-    def __init__(
-        self,
-        dim: int = 1024,
-        num_heads: int = 16,
-        window_size: int = 8,
-        shift_size: int = 4,
-        mlp_ratio: float = 4.0,
-        qkv_bias: bool = True,
-        drop: float = 0.0,
-        attn_drop: float = 0.0,
-        drop_path: float = 0.1,
-        use_relative_position_bias: bool = True,
-        fusion_type: str = "adaptive",  # "weighted", "gated", "adaptive"
-        use_shifted_window: bool = True,
-    ):
-        super().__init__()
-        self.dim = dim
-        self.num_heads = num_heads
-        self.window_size = window_size
-        self.shift_size = shift_size if use_shifted_window else 0
-        self.mlp_ratio = mlp_ratio
-        
-        # =============== Temporal Window Partitioning ===============
-        self.window_partition = TemporalWindowPartition(
-            window_size=window_size,
-            shift_size=self.shift_size,
-        )
-        
-        # =============== Intra-Modal Self-Attention ===============
-        self.norm_audio_self = OmniRMSNorm(dim)
-        self.norm_video_self = OmniRMSNorm(dim)
-        self.norm_image_self = OmniRMSNorm(dim)
-        
-        self.audio_self_attn = WindowCrossAttention(
-            dim=dim,
-            num_heads=num_heads,
-            window_size=window_size,
-            qkv_bias=qkv_bias,
-            attn_drop=attn_drop,
-            proj_drop=drop,
-            use_relative_position_bias=use_relative_position_bias,
-        )
-        
-        self.video_self_attn = WindowCrossAttention(
-            dim=dim,
-            num_heads=num_heads,
-            window_size=window_size,
-            qkv_bias=qkv_bias,
-            attn_drop=attn_drop,
-            proj_drop=drop,
-            use_relative_position_bias=use_relative_position_bias,
-        )
-        
-        self.image_self_attn = WindowCrossAttention(
-            dim=dim,
-            num_heads=num_heads,
-            window_size=window_size,
-            qkv_bias=qkv_bias,
-            attn_drop=attn_drop,
-            proj_drop=drop,
-            use_relative_position_bias=False,  # No temporal bias for static images
-        )
-        
-        # =============== Inter-Modal Cross-Attention ===============
-        # Audio → Video/Image
-        self.norm_audio_cross = OmniRMSNorm(dim)
-        self.audio_to_visual = WindowCrossAttention(
-            dim=dim, num_heads=num_heads, window_size=window_size,
-            qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop,
-        )
-        
-        # Video → Audio/Image
-        self.norm_video_cross = OmniRMSNorm(dim)
-        self.video_to_others = WindowCrossAttention(
-            dim=dim, num_heads=num_heads, window_size=window_size,
-            qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop,
-        )
-        
-        # Image → Audio/Video
-        self.norm_image_cross = OmniRMSNorm(dim)
-        self.image_to_temporal = WindowCrossAttention(
-            dim=dim, num_heads=num_heads, window_size=window_size,
-            qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop,
-        )
-        
-        # =============== Multi-Modal Fusion ===============
-        self.multimodal_fusion = MultiModalFusionLayer(
-            dim=dim,
-            num_modalities=3,
-            fusion_type=fusion_type,
-        )
-        
-        # =============== Feed-Forward Network ===============
-        self.norm_ffn = OmniRMSNorm(dim)
-        mlp_hidden_dim = int(dim * mlp_ratio)
-        self.ffn = nn.Sequential(
-            nn.Linear(dim, mlp_hidden_dim, bias=False),
-            nn.GELU(),
-            nn.Dropout(drop),
-            nn.Linear(mlp_hidden_dim, dim, bias=False),
-            nn.Dropout(drop),
-        )
-        
-        # =============== Stochastic Depth (Drop Path) ===============
-        self.drop_path = nn.Identity() if drop_path <= 0. else nn.Dropout(drop_path)
-        
-        # =============== Output Projections ===============
-        self.output_projection = nn.ModuleDict({
-            'audio': nn.Linear(dim, dim),
-            'video': nn.Linear(dim, dim),
-            'image': nn.Linear(dim, dim),
-        })
-    
-    def forward(
-        self,
-        audio_embeds: Optional[torch.Tensor] = None,  # [B, T_audio, L_audio, D]
-        video_embeds: Optional[torch.Tensor] = None,  # [B, T_video, L_video, D]
-        image_embeds: Optional[torch.Tensor] = None,  # [B, L_image, D]
-        attention_mask: Optional[torch.Tensor] = None,
-        return_intermediates: bool = False,
-    ) -> Dict[str, torch.Tensor]:
-        """
-        Forward pass of the Fancy Multi-Modal Window Attention Block.
-        
-        Args:
-            audio_embeds: Audio embeddings [B, T_audio, L_audio, D]
-                         T_audio: number of audio frames
-                         L_audio: sequence length per frame
-            video_embeds: Video embeddings [B, T_video, L_video, D]
-                         T_video: number of video frames
-                         L_video: sequence length per frame (e.g., patches)
-            image_embeds: Image embeddings [B, L_image, D]
-                         L_image: sequence length (e.g., image patches)
-            attention_mask: Optional attention mask
-            return_intermediates: Whether to return intermediate features
-            
-        Returns:
-            outputs: Dictionary containing processed embeddings for each modality
-                - 'audio': [B, T_audio, L_audio, D]
-                - 'video': [B, T_video, L_video, D]
-                - 'image': [B, L_image, D]
-                - 'fused': [B, L_total, D] (optional)
-        """
-        intermediates = {} if return_intermediates else None
-        
-        # ========== Stage 1: Temporal Window Partitioning ==========
-        partitioned_audio, audio_info = None, None
-        partitioned_video, video_info = None, None
-        
-        if audio_embeds is not None:
-            partitioned_audio, audio_info = self.window_partition.partition(audio_embeds)
-            if return_intermediates:
-                intermediates['audio_windows'] = partitioned_audio
-        
-        if video_embeds is not None:
-            partitioned_video, video_info = self.window_partition.partition(video_embeds)
-            if return_intermediates:
-                intermediates['video_windows'] = partitioned_video
-        
-        # ========== Stage 2: Intra-Modal Self-Attention ==========
-        audio_self_out, video_self_out, image_self_out = None, None, None
-        
-        if audio_embeds is not None:
-            audio_normed = self.norm_audio_self(partitioned_audio)
-            audio_self_out = self.audio_self_attn(audio_normed, audio_normed, audio_normed)
-            audio_self_out = partitioned_audio + self.drop_path(audio_self_out)
-        
-        if video_embeds is not None:
-            video_normed = self.norm_video_self(partitioned_video)
-            video_self_out = self.video_self_attn(video_normed, video_normed, video_normed)
-            video_self_out = partitioned_video + self.drop_path(video_self_out)
-        
-        if image_embeds is not None:
-            image_normed = self.norm_image_self(image_embeds)
-            image_self_out = self.image_self_attn(image_normed, image_normed, image_normed)
-            image_self_out = image_embeds + self.drop_path(image_self_out)
-        
-        # ========== Stage 3: Inter-Modal Cross-Attention ==========
-        audio_cross_out, video_cross_out, image_cross_out = None, None, None
-        
-        # Prepare context (merge windows temporarily for cross-attention)
-        if audio_self_out is not None:
-            audio_merged = self.window_partition.merge(audio_self_out, audio_info)
-        if video_self_out is not None:
-            video_merged = self.window_partition.merge(video_self_out, video_info)
-        
-        # Audio attends to Video and Image
-        if audio_embeds is not None:
-            audio_q = self.norm_audio_cross(audio_merged)
-            
-            # Create key-value context from other modalities
-            kv_list = []
-            if video_embeds is not None:
-                kv_list.append(video_merged)
-            if image_embeds is not None:
-                # Expand image to match temporal dimension
-                B, L_img, D = image_self_out.shape
-                T_audio = audio_merged.shape[1]
-                image_expanded = image_self_out.unsqueeze(1).expand(B, T_audio, L_img, D)
-                kv_list.append(image_expanded)
-            
-            if kv_list:
-                # Concatenate along sequence dimension
-                kv_context = torch.cat([kv.flatten(1, 2) for kv in kv_list], dim=1)
-                kv_context = kv_context.reshape(B, -1, D)
-                
-                audio_cross_out = self.audio_to_visual(
-                    audio_q.flatten(1, 2),
-                    kv_context,
-                    kv_context,
-                    attention_mask
-                )
-                audio_cross_out = audio_cross_out.reshape_as(audio_merged)
-                audio_cross_out = audio_merged + self.drop_path(audio_cross_out)
-            else:
-                audio_cross_out = audio_merged
-        
-        # Video attends to Audio and Image
-        if video_embeds is not None:
-            video_q = self.norm_video_cross(video_merged)
-            
-            kv_list = []
-            if audio_embeds is not None:
-                kv_list.append(audio_merged if audio_cross_out is None else audio_cross_out)
-            if image_embeds is not None:
-                B, L_img, D = image_self_out.shape
-                T_video = video_merged.shape[1]
-                image_expanded = image_self_out.unsqueeze(1).expand(B, T_video, L_img, D)
-                kv_list.append(image_expanded)
-            
-            if kv_list:
-                kv_context = torch.cat([kv.flatten(1, 2) for kv in kv_list], dim=1)
-                kv_context = kv_context.reshape(B, -1, D)
-                
-                video_cross_out = self.video_to_others(
-                    video_q.flatten(1, 2),
-                    kv_context,
-                    kv_context,
-                    attention_mask
-                )
-                video_cross_out = video_cross_out.reshape_as(video_merged)
-                video_cross_out = video_merged + self.drop_path(video_cross_out)
-            else:
-                video_cross_out = video_merged
-        
-        # Image attends to Audio and Video
-        if image_embeds is not None:
-            image_q = self.norm_image_cross(image_self_out)
-            
-            kv_list = []
-            if audio_embeds is not None:
-                # Average pool audio over time for image
-                audio_pooled = (audio_merged if audio_cross_out is None else audio_cross_out).mean(dim=1)
-                kv_list.append(audio_pooled)
-            if video_embeds is not None:
-                # Average pool video over time for image
-                video_pooled = (video_merged if video_cross_out is None else video_cross_out).mean(dim=1)
-                kv_list.append(video_pooled)
-            
-            if kv_list:
-                kv_context = torch.cat(kv_list, dim=1)
-                
-                image_cross_out = self.image_to_temporal(
-                    image_q,
-                    kv_context,
-                    kv_context,
-                    attention_mask
-                )
-                image_cross_out = image_self_out + self.drop_path(image_cross_out)
-            else:
-                image_cross_out = image_self_out
-        
-        # ========== Stage 4: Multi-Modal Fusion ==========
-        # Collect features from all modalities for fusion
-        fusion_features = []
-        if audio_cross_out is not None:
-            audio_flat = audio_cross_out.flatten(1, 2)  # [B, T*L, D]
-            fusion_features.append(audio_flat)
-        if video_cross_out is not None:
-            video_flat = video_cross_out.flatten(1, 2)  # [B, T*L, D]
-            fusion_features.append(video_flat)
-        if image_cross_out is not None:
-            fusion_features.append(image_cross_out)  # [B, L, D]
-        
-        # Pad/align sequence lengths for fusion
-        if len(fusion_features) > 1:
-            max_len = max(f.shape[1] for f in fusion_features)
-            aligned_features = []
-            for feat in fusion_features:
-                if feat.shape[1] < max_len:
-                    pad_len = max_len - feat.shape[1]
-                    feat = F.pad(feat, (0, 0, 0, pad_len))
-                aligned_features.append(feat)
-            
-            # Fuse modalities
-            fused_features = self.multimodal_fusion(aligned_features)
-        else:
-            fused_features = fusion_features[0] if fusion_features else None
-        
-        # ========== Stage 5: Feed-Forward Network ==========
-        if fused_features is not None:
-            fused_normed = self.norm_ffn(fused_features)
-            fused_ffn = self.ffn(fused_normed)
-            fused_features = fused_features + self.drop_path(fused_ffn)
-        
-        # ========== Stage 6: Prepare Outputs ==========
-        outputs = {}
-        
-        # Project back to original shapes
-        if audio_embeds is not None and audio_cross_out is not None:
-            # Partition again for consistency
-            audio_final, _ = self.window_partition.partition(audio_cross_out)
-            audio_final = self.output_projection['audio'](audio_final)
-            audio_final = self.window_partition.merge(audio_final, audio_info)
-            outputs['audio'] = audio_final
-        
-        if video_embeds is not None and video_cross_out is not None:
-            video_final, _ = self.window_partition.partition(video_cross_out)
-            video_final = self.output_projection['video'](video_final)
-            video_final = self.window_partition.merge(video_final, video_info)
-            outputs['video'] = video_final
-        
-        if image_embeds is not None and image_cross_out is not None:
-            image_final = self.output_projection['image'](image_cross_out)
-            outputs['image'] = image_final
-        
-        if fused_features is not None:
-            outputs['fused'] = fused_features
-        
-        if return_intermediates:
-            outputs['intermediates'] = intermediates
-        
-        return outputs
-
-
-# -----------------------------------------------------------------------------
-# 7. Optimization Utilities (FP8, Compilation, Mixed Precision)
-# -----------------------------------------------------------------------------
-
-@dataclass
-class FP8Config:
-    """Configuration for FP8 quantization"""
-    enabled: bool = False
-    margin: int = 0
-    fp8_format: str = "hybrid"  # "e4m3", "e5m2", "hybrid"
-    amax_history_len: int = 1024
-    amax_compute_algo: str = "max"
-
-
-@dataclass  
-class CompilationConfig:
-    """Configuration for torch.compile"""
-    enabled: bool = False
-    mode: str = "reduce-overhead"  # "default", "reduce-overhead", "max-autotune"
-    fullgraph: bool = False
-    dynamic: bool = True
-    backend: str = "inductor"
-
-
-@dataclass
-class MixedPrecisionConfig:
-    """Configuration for mixed precision training/inference"""
-    enabled: bool = True
-    dtype: str = "bfloat16"  # "float16", "bfloat16"
-    use_amp: bool = True
-
-
-class ModelOptimizer:
-    """
-    Unified model optimizer supporting FP8 quantization, torch.compile, 
-    and mixed precision inference.
-    """
-    def __init__(
-        self,
-        fp8_config: Optional[FP8Config] = None,
-        compilation_config: Optional[CompilationConfig] = None,
-        mixed_precision_config: Optional[MixedPrecisionConfig] = None,
-    ):
-        self.fp8_config = fp8_config or FP8Config()
-        self.compilation_config = compilation_config or CompilationConfig()
-        self.mixed_precision_config = mixed_precision_config or MixedPrecisionConfig()
-        
-        # Setup mixed precision
-        self._setup_mixed_precision()
-    
-    def _setup_mixed_precision(self):
-        """Setup mixed precision context"""
-        if self.mixed_precision_config.enabled:
-            dtype_map = {
-                "float16": torch.float16,
-                "bfloat16": torch.bfloat16,
-            }
-            self.dtype = dtype_map.get(self.mixed_precision_config.dtype, torch.bfloat16)
-        else:
-            self.dtype = torch.float32
-    
-    @contextmanager
-    def autocast_context(self):
-        """Context manager for automatic mixed precision"""
-        if self.mixed_precision_config.enabled and self.mixed_precision_config.use_amp:
-            with torch.autocast(device_type='cuda', dtype=self.dtype):
-                yield
-        else:
-            yield
-    
-    def _compile_model(self, model: nn.Module) -> nn.Module:
-        """Compile model using torch.compile"""
-        if not self.compilation_config.enabled or not HAS_TORCH_COMPILE:
-            return model
-        
-        return torch.compile(
-            model,
-            mode=self.compilation_config.mode,
-            fullgraph=self.compilation_config.fullgraph,
-            dynamic=self.compilation_config.dynamic,
-            backend=self.compilation_config.backend,
-        )
-    
-    def _quantize_model_fp8(self, model: nn.Module) -> nn.Module:
-        """Apply FP8 quantization using Transformer Engine"""
-        if not self.fp8_config.enabled or not HAS_TRANSFORMER_ENGINE:
-            return model
-        
-        # Convert compatible layers to FP8
-        for name, module in model.named_modules():
-            if isinstance(module, nn.Linear):
-                # Replace with TE FP8 Linear
-                fp8_linear = te.Linear(
-                    module.in_features,
-                    module.out_features,
-                    bias=module.bias is not None,
-                )
-                # Copy weights
-                fp8_linear.weight.data.copy_(module.weight.data)
-                if module.bias is not None:
-                    fp8_linear.bias.data.copy_(module.bias.data)
-                
-                # Replace module
-                parent_name = '.'.join(name.split('.')[:-1])
-                child_name = name.split('.')[-1]
-                if parent_name:
-                    parent = dict(model.named_modules())[parent_name]
-                    setattr(parent, child_name, fp8_linear)
-        
-        return model
-    
-    def optimize_model(
-        self,
-        model: nn.Module,
-        apply_compilation: bool = True,
-        apply_quantization: bool = True,
-        apply_mixed_precision: bool = True,
-    ) -> nn.Module:
-        """
-        Apply all optimizations to model.
-        
-        Args:
-            model: Model to optimize
-            apply_compilation: Whether to compile with torch.compile
-            apply_quantization: Whether to apply FP8 quantization
-            apply_mixed_precision: Whether to convert to mixed precision dtype
-            
-        Returns:
-            Optimized model
-        """
-        # Apply FP8 quantization first
-        if apply_quantization and self.fp8_config.enabled:
-            model = self._quantize_model_fp8(model)
-        
-        # Convert to mixed precision dtype
-        if apply_mixed_precision and self.mixed_precision_config.enabled:
-            model = model.to(dtype=self.dtype)
-        
-        # Compile model last
-        if apply_compilation and self.compilation_config.enabled:
-            model = self._compile_model(model)
-        
-        return model
-
-
-@contextmanager
-def optimized_inference_mode(
-    enable_cudnn_benchmark: bool = True,
-    enable_tf32: bool = True,
-    enable_flash_sdp: bool = True,
-):
-    """
-    Context manager for optimized inference with various PyTorch optimizations.
-    
-    Args:
-        enable_cudnn_benchmark: Enable cuDNN autotuner
-        enable_tf32: Enable TF32 for faster matmul on Ampere+ GPUs
-        enable_flash_sdp: Enable Flash Attention in scaled_dot_product_attention
-    """
-    # Save original states
-    orig_benchmark = torch.backends.cudnn.benchmark
-    orig_tf32_matmul = torch.backends.cuda.matmul.allow_tf32
-    orig_tf32_cudnn = torch.backends.cudnn.allow_tf32
-    orig_sdp_flash = torch.backends.cuda.flash_sdp_enabled()
-    
-    try:
-        # Enable optimizations
-        torch.backends.cudnn.benchmark = enable_cudnn_benchmark
-        torch.backends.cuda.matmul.allow_tf32 = enable_tf32
-        torch.backends.cudnn.allow_tf32 = enable_tf32
-        
-        if enable_flash_sdp:
-            torch.backends.cuda.enable_flash_sdp(True)
-        
-        yield
-        
-    finally:
-        # Restore original states
-        torch.backends.cudnn.benchmark = orig_benchmark
-        torch.backends.cuda.matmul.allow_tf32 = orig_tf32_matmul
-        torch.backends.cudnn.allow_tf32 = orig_tf32_cudnn
-        torch.backends.cuda.enable_flash_sdp(orig_sdp_flash)

push.sh

@@ -0,0 +1,13 @@
+
+# 设置仓库级别用户名
+git config user.name "selfitcamera"
+git config user.email "ethan.blake@heybeauty.ai"
+
+# 验证
+git config user.name
+git config user.email
+
+
+git add .
+git commit -m "init"
+git push

util.py

@@ -0,0 +1,729 @@
+
+import os
+import sys
+import cv2
+import json
+import random
+import time
+import datetime
+import requests
+import func_timeout
+import numpy as np
+import gradio as gr
+import boto3
+import tempfile
+import io
+import uuid
+from botocore.client import Config
+from PIL import Image
+
+
+# TOKEN = os.environ['TOKEN']
+# APIKEY = os.environ['APIKEY']
+# UKAPIURL = os.environ['UKAPIURL']
+
+OneKey = os.environ['OneKey'].strip()
+OneKey = OneKey.split("#")
+TOKEN = OneKey[0]
+APIKEY = OneKey[1]
+UKAPIURL = OneKey[2]
+LLMKEY = OneKey[3]
+R2_ACCESS_KEY = OneKey[4]
+R2_SECRET_KEY = OneKey[5]
+R2_ENDPOINT = OneKey[6]
+
+
+# tmpFolder is no longer needed since we upload directly from memory
+# tmpFolder = "tmp"
+# os.makedirs(tmpFolder, exist_ok=True)
+
+
+# Legacy function - no longer used since we upload directly from memory
+# def upload_user_img(clientIp, timeId, img):
+#     fileName = clientIp.replace(".", "")+str(timeId)+".jpg"
+#     local_path = os.path.join(tmpFolder, fileName)
+#     img = cv2.imread(img)
+#     cv2.imwrite(os.path.join(tmpFolder, fileName), img)
+# 
+#     json_data = {
+#         "token": TOKEN,  
+#         "input1": fileName,
+#         "input2": "",
+#         "protocol": "",
+#         "cloud": "ali"
+#     }
+#     
+#     session = requests.session()
+#     ret = requests.post(
+#         f"{UKAPIURL}/upload",
+#         headers={'Content-Type': 'application/json'},
+#         json=json_data
+#     )
+#     
+#     res = ""
+#     if ret.status_code==200:
+#         if 'upload1' in ret.json():
+#             upload_url = ret.json()['upload1']
+#             headers = {'Content-Type': 'image/jpeg'}
+#             response = session.put(upload_url, data=open(local_path, 'rb').read(), headers=headers)
+#             # print(response.status_code)
+#             if response.status_code == 200:
+#                 res = upload_url
+#     if os.path.exists(local_path):
+#         os.remove(local_path)
+#     return res
+
+
+class R2Api:
+
+    def __init__(self, session=None):
+        super().__init__()
+        self.R2_BUCKET = "omni-creator"
+        self.domain = "https://www.omnicreator.net/"
+        self.R2_ACCESS_KEY = R2_ACCESS_KEY
+        self.R2_SECRET_KEY = R2_SECRET_KEY
+        self.R2_ENDPOINT = R2_ENDPOINT
+
+        self.client = boto3.client(
+            "s3",
+            endpoint_url=self.R2_ENDPOINT,
+            aws_access_key_id=self.R2_ACCESS_KEY,
+            aws_secret_access_key=self.R2_SECRET_KEY,
+            config=Config(signature_version="s3v4")
+        )
+
+        self.session = requests.Session() if session is None else session
+
+    def upload_from_memory(self, image_data, filename, content_type='image/jpeg'):
+        """
+        Upload image data directly from memory to R2
+        
+        Args:
+            image_data (bytes): Image data in bytes
+            filename (str): Filename for the uploaded file
+            content_type (str): MIME type of the image
+            
+        Returns:
+            str: URL of the uploaded file
+        """
+        t1 = time.time()
+        headers = {"Content-Type": content_type}
+        
+        cloud_path = f"ImageEdit/Uploads/{str(datetime.date.today())}/{filename}"
+        url = self.client.generate_presigned_url(
+            "put_object",
+            Params={"Bucket": self.R2_BUCKET, "Key": cloud_path, "ContentType": content_type},
+            ExpiresIn=604800
+        )
+
+        retry_count = 0
+        while retry_count < 3:
+            try:
+                response = self.session.put(url, data=image_data, headers=headers, timeout=15)
+                if response.status_code == 200:
+                    break
+                else:
+                    print(f"⚠️ Upload failed with status code: {response.status_code}")
+                    retry_count += 1
+            except (requests.exceptions.Timeout, requests.exceptions.RequestException) as e:
+                print(f"⚠️ Upload retry {retry_count + 1}/3 failed: {e}")
+                retry_count += 1
+                if retry_count == 3:
+                    raise Exception(f'Failed to upload file to R2 after 3 retries! Last error: {str(e)}')
+                time.sleep(1)  # 等待1秒后重试
+                continue
+        print("upload_from_memory time is ====>", time.time() - t1)
+        return f"{self.domain}{cloud_path}"
+
+def upload_user_img_r2(clientIp, timeId, pil_image):
+    """
+    Upload PIL Image directly to R2 without saving to local file
+    
+    Args:
+        clientIp (str): Client IP address
+        timeId (int): Timestamp
+        pil_image (PIL.Image): PIL Image object
+        
+    Returns:
+        str: Uploaded URL
+    """
+    # Generate unique filename using UUID to prevent file conflicts in concurrent environment
+    unique_id = str(uuid.uuid4())
+    fileName = f"user_img_{unique_id}_{timeId}.jpg"
+    
+    # Convert PIL Image to bytes
+    img_buffer = io.BytesIO()
+    if pil_image.mode != 'RGB':
+        pil_image = pil_image.convert('RGB')
+    pil_image.save(img_buffer, format='JPEG', quality=95)
+    img_data = img_buffer.getvalue()
+    
+    # Upload directly from memory
+    res = R2Api().upload_from_memory(img_data, fileName, 'image/jpeg')
+    return res
+
+
+
+def create_mask_from_layers(base_image, layers):
+    """
+    Create mask image from ImageEditor layers
+    
+    Args:
+        base_image (PIL.Image): Original image
+        layers (list): ImageEditor layer data
+        
+    Returns:
+        PIL.Image: Black and white mask image
+    """
+    from PIL import Image, ImageDraw
+    import numpy as np
+    
+    # Create blank mask with same size as original image
+    mask = Image.new('L', base_image.size, 0)  # 'L' mode is grayscale, 0 is black
+    
+    if not layers:
+        return mask
+    
+    # Iterate through all layers, set drawn areas to white
+    for layer in layers:
+        if layer is not None:
+            # Convert layer to numpy array
+            layer_array = np.array(layer)
+            
+            # Check layer format
+            if len(layer_array.shape) == 3:  # RGB/RGBA format
+                # If RGBA, check alpha channel
+                if layer_array.shape[2] == 4:
+                    # Use alpha channel as mask
+                    alpha_channel = layer_array[:, :, 3]
+                    # Set non-transparent areas (alpha > 0) to white
+                    mask_array = np.where(alpha_channel > 0, 255, 0).astype(np.uint8)
+                else:
+                    # RGB format, check if not pure black (0,0,0)
+                    # Assume drawn areas are non-black
+                    non_black = np.any(layer_array > 0, axis=2)
+                    mask_array = np.where(non_black, 255, 0).astype(np.uint8)
+            elif len(layer_array.shape) == 2:  # Grayscale
+                # Use grayscale values directly, set non-zero areas to white
+                mask_array = np.where(layer_array > 0, 255, 0).astype(np.uint8)
+            else:
+                continue
+            
+            # Convert mask_array to PIL image and merge into total mask
+            layer_mask = Image.fromarray(mask_array, mode='L')
+            # Resize to match original image
+            if layer_mask.size != base_image.size:
+                layer_mask = layer_mask.resize(base_image.size, Image.LANCZOS)
+            
+            # Merge masks (use maximum value to ensure all drawn areas are included)
+            mask_array_current = np.array(mask)
+            layer_mask_array = np.array(layer_mask)
+            combined_mask_array = np.maximum(mask_array_current, layer_mask_array)
+            mask = Image.fromarray(combined_mask_array, mode='L')
+    
+    return mask
+
+
+def upload_mask_image_r2(client_ip, time_id, mask_image):
+    """
+    Upload mask image to R2 directly from memory
+    
+    Args:
+        client_ip (str): Client IP
+        time_id (int): Timestamp
+        mask_image (PIL.Image): Mask image
+        
+    Returns:
+        str: Uploaded URL
+    """
+    # Generate unique filename using UUID to prevent file conflicts in concurrent environment
+    unique_id = str(uuid.uuid4())
+    file_name = f"mask_img_{unique_id}_{time_id}.png"
+    
+    try:
+        # Convert mask image to bytes
+        img_buffer = io.BytesIO()
+        mask_image.save(img_buffer, format='PNG')
+        img_data = img_buffer.getvalue()
+        
+        # Upload directly from memory
+        res = R2Api().upload_from_memory(img_data, file_name, 'image/png')
+        
+        return res
+    except Exception as e:
+        print(f"Failed to upload mask image: {e}")
+        return None
+
+
+
+def submit_image_edit_task(user_image_url, prompt, task_type="80", mask_image_url="", reference_image_url=""):
+    """
+    Submit image editing task with improved error handling using API v2
+    """
+    headers = {
+        'Content-Type': 'application/json',
+        'Authorization': f'Bearer {APIKEY}'
+    }
+    
+    data = {
+        "user_image": user_image_url,
+        "user_mask": mask_image_url,
+        "type": task_type,
+        "text": prompt,
+        "user_uuid": APIKEY,
+        "priority": 0,
+        "secret_key": "219ngu"
+    }
+
+    if reference_image_url:
+        data["user_image2"] = reference_image_url
+    
+    retry_count = 0
+    max_retries = 3
+    
+    while retry_count < max_retries:
+        try:
+            response = requests.post(
+                f'{UKAPIURL}/public_image_edit_v2',
+                headers=headers,
+                json=data,
+                timeout=30  # 增加超时时间
+            )
+            
+            if response.status_code == 200:
+                result = response.json()
+                if result.get('code') == 0:
+                    return result['data']['task_id'], None
+                else:
+                    return None, f"API Error: {result.get('message', 'Unknown error')}"
+            elif response.status_code in [502, 503, 504]:  # 服务器错误，可以重试
+                retry_count += 1
+                if retry_count < max_retries:
+                    print(f"⚠️ Server error {response.status_code}, retrying {retry_count}/{max_retries}")
+                    time.sleep(2)  # 等待2秒后重试
+                    continue
+                else:
+                    return None, f"HTTP Error after {max_retries} retries: {response.status_code}"
+            else:
+                return None, f"HTTP Error: {response.status_code}"
+                
+        except (requests.exceptions.Timeout, requests.exceptions.ConnectionError) as e:
+            retry_count += 1
+            if retry_count < max_retries:
+                print(f"⚠️ Network error, retrying {retry_count}/{max_retries}: {e}")
+                time.sleep(2)
+                continue
+            else:
+                return None, f"Network error after {max_retries} retries: {str(e)}"
+        except Exception as e:
+            return None, f"Request Exception: {str(e)}"
+    
+    return None, f"Failed after {max_retries} retries"
+
+
+def check_task_status(task_id):
+    """
+    Query task status with improved error handling using API v2
+    """
+    headers = {
+        'Content-Type': 'application/json',
+        'Authorization': f'Bearer {APIKEY}'
+    }
+    
+    data = {
+        "task_id": task_id
+    }
+    
+    retry_count = 0
+    max_retries = 2  # 状态查询重试次数少一些
+    
+    while retry_count < max_retries:
+        try:
+            response = requests.post(
+                f'{UKAPIURL}/status_image_edit_v2',
+                headers=headers,
+                json=data,
+                timeout=15  # 状态查询超时时间短一些
+            )
+            
+            if response.status_code == 200:
+                result = response.json()
+                if result.get('code') == 0:
+                    task_data = result['data']
+                    status = task_data['status']
+                    image_url = task_data.get('image_url')
+                    
+                    # Extract and log queue information for better user feedback
+                    queue_info = task_data.get('queue_info', {})
+                    if queue_info:
+                        tasks_ahead = queue_info.get('tasks_ahead', 0)
+                        current_priority = queue_info.get('current_priority', 0)
+                        description = queue_info.get('description', '')
+                        print(f"📊 Queue Status - Tasks ahead: {tasks_ahead}, Priority: {current_priority}, Status: {status}")
+                    
+                    return status, image_url, task_data
+                else:
+                    return 'error', None, result.get('message', 'Unknown error')
+            elif response.status_code in [502, 503, 504]:  # 服务器错误，可以重试
+                retry_count += 1
+                if retry_count < max_retries:
+                    print(f"⚠️ Status check server error {response.status_code}, retrying {retry_count}/{max_retries}")
+                    time.sleep(1)  # 状态查询重试间隔短一些
+                    continue
+                else:
+                    return 'error', None, f"HTTP Error after {max_retries} retries: {response.status_code}"
+            else:
+                return 'error', None, f"HTTP Error: {response.status_code}"
+                
+        except (requests.exceptions.Timeout, requests.exceptions.ConnectionError) as e:
+            retry_count += 1
+            if retry_count < max_retries:
+                print(f"⚠️ Status check network error, retrying {retry_count}/{max_retries}: {e}")
+                time.sleep(1)
+                continue
+            else:
+                return 'error', None, f"Network error after {max_retries} retries: {str(e)}"
+        except Exception as e:
+            return 'error', None, f"Request Exception: {str(e)}"
+    
+    return 'error', None, f"Failed after {max_retries} retries"
+
+
+def process_image_edit(img_input, prompt, reference_image=None, progress_callback=None):
+    """
+    Complete process for image editing
+    
+    Args:
+        img_input: Can be file path (str) or PIL Image object
+        prompt: Editing instructions
+        progress_callback: Progress callback function
+    """
+    try:
+        # Generate client IP and timestamp
+        client_ip = "127.0.0.1"  # Default IP
+        time_id = int(time.time())
+        
+        # Process input image - supports PIL Image and file path
+        if hasattr(img_input, 'save'):  # PIL Image object
+            pil_image = img_input
+            print(f"💾 Using PIL Image directly from memory")
+        else:
+            # Load from file path
+            pil_image = Image.open(img_input)
+            print(f"📁 Loaded image from file: {img_input}")
+        
+        if progress_callback:
+            progress_callback("uploading image...")
+        
+        # Upload user image directly from memory
+        uploaded_url = upload_user_img_r2(client_ip, time_id, pil_image)
+        if not uploaded_url:
+            return None, "image upload failed", None
+        
+        # Extract actual image URL from upload URL
+        if "?" in uploaded_url:
+            uploaded_url = uploaded_url.split("?")[0]
+        
+        if progress_callback:
+            progress_callback("submitting edit task...")
+        
+        reference_url = ""
+        if reference_image is not None:
+            try:
+                if progress_callback:
+                    progress_callback("uploading reference image...")
+
+                if hasattr(reference_image, 'save'):
+                    reference_pil = reference_image
+                else:
+                    reference_pil = Image.open(reference_image)
+
+                reference_url = upload_user_img_r2(client_ip, time_id, reference_pil)
+                if not reference_url:
+                    return None, "reference image upload failed", None
+
+                if "?" in reference_url:
+                    reference_url = reference_url.split("?")[0]
+            except Exception as e:
+                return None, f"reference image processing failed: {str(e)}", None
+
+        # Submit image editing task
+        task_id, error = submit_image_edit_task(uploaded_url, prompt, reference_image_url=reference_url)
+        if error:
+            return None, error, None
+        
+        if progress_callback:
+            progress_callback(f"task submitted, ID: {task_id}, processing...")
+        
+        # Wait for task completion
+        max_attempts = 60  # Wait up to 10 minutes
+        task_uuid = None
+        for attempt in range(max_attempts):
+            status, output_url, task_data = check_task_status(task_id)
+            
+            # Extract task_uuid from task_data
+            if task_data and isinstance(task_data, dict):
+                task_uuid = task_data.get('uuid', None)
+            
+            if status == 'completed':
+                if output_url:
+                    return output_url, "image edit completed", task_uuid
+                else:
+                    return None, "Task completed but no result image returned", task_uuid
+            elif status == 'error' or status == 'failed':
+                return None, f"task processing failed: {task_data}", task_uuid
+            elif status in ['queued', 'processing', 'running', 'created', 'working']:
+                # Enhanced progress message with queue info and website promotion
+                if progress_callback and task_data and isinstance(task_data, dict):
+                    queue_info = task_data.get('queue_info', {})
+                    if queue_info and status in ['queued', 'created']:
+                        tasks_ahead = queue_info.get('tasks_ahead', 0)
+                        current_priority = queue_info.get('current_priority', 0)
+                        if tasks_ahead > 0:
+                            progress_callback(f"⏳ Queue: {tasks_ahead} tasks ahead | Low priority | Visit website for instant processing → https://omnicreator.net/#generator")
+                        else:
+                            progress_callback(f"🚀 Processing your image editing request...")
+                    elif status == 'processing':
+                        progress_callback(f"🎨 AI is processing... Please wait")
+                    elif status in ['running', 'working']:
+                        progress_callback(f"⚡ Generating... Almost done")
+                    else:
+                        progress_callback(f"📋 Task status: {status}")
+                else:
+                    if progress_callback:
+                        progress_callback(f"task processing... (status: {status})")
+                time.sleep(1)
+            else:
+                if progress_callback:
+                    progress_callback(f"unknown status: {status}")
+                time.sleep(1)
+        
+        return None, "task processing timeout", task_uuid
+        
+    except Exception as e:
+        return None, f"error occurred during processing: {str(e)}", None
+
+
+def process_local_image_edit(base_image, layers, prompt, reference_image=None, progress_callback=None, use_example_mask=None):
+    """
+    处理局部图片编辑的完整流程
+    
+    Args:
+        base_image (PIL.Image): 原始图片
+        layers (list): ImageEditor的层数据
+        prompt (str): 编辑指令
+        progress_callback: 进度回调函数
+    """
+    try:
+        # Generate client IP and timestamp
+        client_ip = "127.0.0.1"  # Default IP
+        time_id = int(time.time())
+        
+        if progress_callback:
+            progress_callback("creating mask image...")
+        
+        # Check if we should use example mask (backdoor for example case)
+        if use_example_mask:
+            # Load local mask file for example
+            try:
+                from PIL import Image
+                import os
+                
+                # Check if base_image is valid
+                if base_image is None:
+                    return None, "Base image is None, cannot process example mask", None
+                
+                if os.path.exists(use_example_mask):
+                    mask_image = Image.open(use_example_mask)
+                    
+                    # Ensure mask has same size as base image
+                    if hasattr(base_image, 'size') and mask_image.size != base_image.size:
+                        mask_image = mask_image.resize(base_image.size)
+                    
+                    # Ensure mask is in L mode (grayscale)
+                    if mask_image.mode != 'L':
+                        mask_image = mask_image.convert('L')
+                    
+                    print(f"🎭 Using example mask from: {use_example_mask}, size: {mask_image.size}")
+                else:
+                    return None, f"Example mask file not found: {use_example_mask}", None
+            except Exception as e:
+                import traceback
+                traceback.print_exc()
+                return None, f"Failed to load example mask: {str(e)}", None
+        else:
+            # Normal case: create mask from layers
+            mask_image = create_mask_from_layers(base_image, layers)
+            
+            # 检查mask是否有内容
+            mask_array = np.array(mask_image)
+            if np.max(mask_array) == 0:
+                return None, "please draw mask", None
+        
+        # Print mask statistics
+        if not use_example_mask:
+            print(f"📝 创建mask图片成功，绘制区域像素数: {np.sum(mask_array > 0)}")
+        else:
+            mask_array = np.array(mask_image)
+            print(f"🎭 Example mask loaded successfully, mask pixels: {np.sum(mask_array > 0)}")
+        
+        if progress_callback:
+            progress_callback("uploading original image...")
+        
+        # 直接从内存上传原始图片
+        uploaded_url = upload_user_img_r2(client_ip, time_id, base_image)
+        if not uploaded_url:
+            return None, "original image upload failed", None
+        
+        # 从上传 URL 中提取实际的图片 URL
+        if "?" in uploaded_url:
+            uploaded_url = uploaded_url.split("?")[0]
+        
+        if progress_callback:
+            progress_callback("uploading mask image...")
+        
+        # 直接从内存上传mask图片
+        mask_url = upload_mask_image_r2(client_ip, time_id, mask_image)
+        if not mask_url:
+            return None, "mask image upload failed", None
+        
+        # 从上传 URL 中提取实际的图片 URL
+        if "?" in mask_url:
+            mask_url = mask_url.split("?")[0]
+        
+        reference_url = ""
+        if reference_image is not None:
+            try:
+                if progress_callback:
+                    progress_callback("uploading reference image...")
+
+                if hasattr(reference_image, 'save'):
+                    reference_pil = reference_image
+                else:
+                    reference_pil = Image.open(reference_image)
+
+                reference_url = upload_user_img_r2(client_ip, time_id, reference_pil)
+                if not reference_url:
+                    return None, "reference image upload failed", None
+
+                if "?" in reference_url:
+                    reference_url = reference_url.split("?")[0]
+            except Exception as e:
+                return None, f"reference image processing failed: {str(e)}", None
+
+        print(f"📤 图片上传成功:")
+        print(f"   原始图片: {uploaded_url}")
+        print(f"   Mask图片: {mask_url}")
+        if reference_url:
+            print(f"   参考图片: {reference_url}")
+        
+        if progress_callback:
+            progress_callback("submitting local edit task...")
+        
+        # 提交局部图片编辑任务 (task_type=81)
+        task_id, error = submit_image_edit_task(
+            uploaded_url,
+            prompt,
+            task_type="81",
+            mask_image_url=mask_url,
+            reference_image_url=reference_url
+        )
+        if error:
+            return None, error, None
+        
+        if progress_callback:
+            progress_callback(f"task submitted, ID: {task_id}, processing...")
+        
+        print(f"🚀 局部编辑任务已提交，任务ID: {task_id}")
+        
+        # Wait for task completion
+        max_attempts = 60  # Wait up to 10 minutes
+        task_uuid = None
+        for attempt in range(max_attempts):
+            status, output_url, task_data = check_task_status(task_id)
+            
+            # Extract task_uuid from task_data
+            if task_data and isinstance(task_data, dict):
+                task_uuid = task_data.get('uuid', None)
+            
+            if status == 'completed':
+                if output_url:
+                    print(f"✅ 局部编辑任务完成，结果: {output_url}")
+                    return output_url, "local image edit completed", task_uuid
+                else:
+                    return None, "task completed but no result image returned", task_uuid
+            elif status == 'error' or status == 'failed':
+                return None, f"task processing failed: {task_data}", task_uuid
+            elif status in ['queued', 'processing', 'running', 'created', 'working']:
+                # Enhanced progress message with queue info and website promotion
+                if progress_callback and task_data and isinstance(task_data, dict):
+                    queue_info = task_data.get('queue_info', {})
+                    if queue_info and status in ['queued', 'created']:
+                        tasks_ahead = queue_info.get('tasks_ahead', 0)
+                        current_priority = queue_info.get('current_priority', 0)
+                        if tasks_ahead > 0:
+                            progress_callback(f"⏳ Queue: {tasks_ahead} tasks ahead | Low priority | Visit website for instant processing → https://omnicreator.net/#generator")
+                        else:
+                            progress_callback(f"🚀 Processing your local editing request...")
+                    elif status == 'processing':
+                        progress_callback(f"🎨 AI is processing... Please wait")
+                    elif status in ['running', 'working']:
+                        progress_callback(f"⚡ Generating... Almost done")
+                    else:
+                        progress_callback(f"📋 Task status: {status}")
+                else:
+                    if progress_callback:
+                        progress_callback(f"processing... (status: {status})")
+                time.sleep(1)  # Wait 1 second before retry
+            else:
+                if progress_callback:
+                    progress_callback(f"unknown status: {status}")
+                time.sleep(1)
+        
+        return None, "task processing timeout", task_uuid
+        
+    except Exception as e:
+        print(f"❌ 局部编辑处理异常: {str(e)}")
+        return None, f"error occurred during processing: {str(e)}", None
+
+
+def download_and_check_result_nsfw(image_url, nsfw_detector=None):
+    """
+    下载结果图片并进行NSFW检测
+    
+    Args:
+        image_url (str): 结果图片URL
+        nsfw_detector: NSFW检测器实例
+        
+    Returns:
+        tuple: (is_nsfw, error_message)
+    """
+    if nsfw_detector is None:
+        return False, None
+    
+    try:
+        # 下载图片
+        response = requests.get(image_url, timeout=30)
+        if response.status_code != 200:
+            return False, f"Failed to download result image: HTTP {response.status_code}"
+        
+        # 将图片数据转换为PIL Image
+        image_data = io.BytesIO(response.content)
+        result_image = Image.open(image_data)
+        
+        # 进行NSFW检测
+        nsfw_result = nsfw_detector.predict_pil_label_only(result_image)
+        
+        is_nsfw = nsfw_result.lower() == "nsfw"
+        print(f"🔍 结果图片NSFW检测: {'❌❌❌ ' + nsfw_result if is_nsfw else '✅✅✅ ' + nsfw_result}")
+        
+        return is_nsfw, None
+        
+    except Exception as e:
+        print(f"⚠️ 结果图片NSFW检测失败: {e}")
+        return False, f"Failed to check result image: {str(e)}"
+
+
+if __name__ == "__main__":
+    
+    pass