rajeshrai577/z-image-turbo-lora-nepali

Nepali Style – Z-Image Turbo LoRA (Testing Only)

IMPORTANT
This LoRA is released ONLY FOR TESTING AND EXPERIMENTAL PURPOSES.
It is not intended for production use.

---

Overview

This repository contains a style LoRA trained on Z-Image Turbo using AI Toolkit.
The goal of this project is to experiment and test LoRA fine-tuning workflows on the Z-Image Turbo architecture with a very small dataset.

The LoRA introduces a Nepali-inspired visual style while maintaining the speed and low-step inference behavior of Z-Image Turbo.

---

Testing Purpose Disclaimer

This model is for testing only

• Trained on a very small dataset (10 images)
• Results may be unstable or inconsistent
• Not optimized for real-world or commercial deployment
• Shared for research, learning, and experimentation

Use at your own risk.

---

Trigger Word

There is NO mandatory trigger word.

Optionally, you may use: nepali_style to help reinforce the learned style.

---

Recommended Inference Settings

Z-Image Turbo is extremely sensitive.
Stay within these limits for best results:

• LoRA strength: 0.6 – 0.9
• Inference steps: 4 – 8
• CFG / Guidance scale: 1 – 2
• Sampler: FlowMatch or Euler (Turbo-compatible)

Avoid LoRA strength above 1.0 — this may degrade output quality.

---

Example Prompts

• a cinematic portrait of a nepali man, traditional clothing, soft natural lighting
• a woman in nepali_style, studio lighting, ultra detailed portrait

Training Details

• Base model: Tongyi-MAI/Z-Image-Turbo
• Architecture: Z-Image Turbo
• Training tool: AI Toolkit (diffusion_trainer)
• Training steps: 3000
• Batch size: 1
• Resolution: 512 × 512
• Dataset size: 10 images
• Style/content balance: Balanced
• Noise scheduler: FlowMatch
• Loss type: MSE
• Precision: BF16
• Optimizer: AdamW (8-bit)
• Learning rate: 1e-4
• UNet: Trained
• Text encoder: Frozen (not trained)
• Latent caching: Enabled
• Text embedding caching: Enabled
• Quantization: QFloat8 (model + text encoder)
• Low VRAM mode: Enabled
• Training adapter:
  ostris/zimage_turbo_training_adapter_v2

LoRA Configuration

• Linear rank: 32 (alpha 32)
• Conv rank: 16 (alpha 16)
• LoKr: Full-rank enabled

---

Samples

Sample images included in this repository were generated using:

• Steps: 8
• CFG: 1
• Sampler: FlowMatch
• Seed: 42

(See images displayed below.)

---

Usage (Diffusers)

from diffusers import AutoPipelineForText2Image
import torch

pipe = AutoPipelineForText2Image.from_pretrained(
    "Tongyi-MAI/Z-Image-Turbo",
    torch_dtype=torch.float16
)

pipe.load_lora_weights("your-username/nepali-style-zimage-turbo-lora")
pipe.to("cuda")

image = pipe(
    "a cinematic portrait of a nepali man",
    num_inference_steps=6,
    guidance_scale=1.2
).images[0]

image.save("output.png")

---

Hugging Face Space

A demo Space may be provided to test this LoRA interactively.
This Space is also intended only for testing and experimentation.

---

License

This model is released under the OpenRAIL++ license.

• This LoRA is a derivative work of Z-Image Turbo
• All base model license terms apply
• Users are responsible for generated content

---

Final Disclaimer

• This project is experimental
• Outputs are not guaranteed
• The author is not responsible for misuse
• Please avoid generating illegal, harmful, or copyrighted content

---

Credits

• Base model: Tongyi-MAI / Z-Image Turbo
• Training adapter: Ostris
• LoRA training & testing: @rajeshrai577