Add Adaptive Strength Diffusion

app.py

@@ -186,7 +186,7 @@ pipe = ObjectClearPipeline.from_pretrained_with_custom_modules(
 pipe.to(device)
     
 @spaces.GPU
-def process(image_state, interactive_state, mask_dropdown, guidance_scale, seed, num_inference_steps, strength         
+def process(image_state, interactive_state, mask_dropdown, guidance_scale, seed, num_inference_steps         
             ):
     generator = torch.Generator(device="cuda").manual_seed(seed)
     image_np = image_state["origin_image"]
@@ -219,7 +219,6 @@ def process(image_state, interactive_state, mask_dropdown, guidance_scale, seed,
         mask_image=mask,
         generator=generator,
         num_inference_steps=num_inference_steps,
-        strength=strength,
         guidance_scale=guidance_scale,
         height=h,
         width=w,
@@ -432,13 +431,6 @@ with gr.Blocks(css=custom_css) as demo:
             )
             
             with gr.Accordion('ObjectClear Settings', open=True):
-                strength = gr.Radio(
-                    choices=[0.99, 1.0],
-                    value=0.99,
-                    label="Strength",
-                    info="0.99 better preserves the background and color; use 1.0 if object/shadow is not fully removed (default: 0.99)"
-                )
-                
                 guidance_scale = gr.Slider(
                     minimum=1, maximum=10, step=0.5, value=2.5,
                     label="Guidance Scale",
@@ -517,8 +509,7 @@ with gr.Blocks(css=custom_css) as demo:
             mask_dropdown,
             guidance_scale,
             seed,
-            num_inference_steps,
-            strength
+            num_inference_steps
         ],
         outputs=[
             output_image_component, output_compare_image_component

pipeline_objectclear.py

@@ -1352,7 +1352,7 @@ class ObjectClearPipeline(
         height: Optional[int] = None,
         width: Optional[int] = None,
         padding_mask_crop: Optional[int] = None,
-        strength: float = 0.9999,
+        strength: float = 1.0
         num_inference_steps: int = 50,
         timesteps: List[int] = None,
         sigmas: List[float] = None,
@@ -1426,7 +1426,7 @@ class ObjectClearPipeline(
                 on `padding_mask_crop`. The image and mask_image will then be cropped based on the expanded area before
                 resizing to the original image size for inpainting. This is useful when the masked area is small while
                 the image is large and contain information irrelevant for inpainting, such as background.
-            strength (`float`, *optional*, defaults to 0.9999):
+            strength (`float`, *optional*, defaults to 1.0):
                 Conceptually, indicates how much to transform the masked portion of the reference `image`. Must be
                 between 0 and 1. `image` will be used as a starting point, adding more noise to it the larger the
                 `strength`. The number of denoising steps depends on the amount of noise initially added. When
@@ -1914,6 +1914,17 @@ class ObjectClearPipeline(
                 # progressive attention mask blending
                 fuse_index = 5
                 if self.config.apply_attention_guided_fusion:
+                    if i == 0:
+                        init_latents_proper = image_latents
+                        init_mask = mask[0:1]
+
+                        noise_timestep = timesteps[i + 1]
+                        init_latents_proper = self.scheduler.add_noise(
+                            init_latents_proper, noise, torch.tensor([noise_timestep])
+                        )
+                        
+                        latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
                     if i == len(timesteps) - 1:
                         attn_key, attn_map = next(iter(self.cross_attention_scores.items()))
                         attn_map = self.resize_attn_map_divide2(attn_map, mask, fuse_index)