Z-Image-Turbo-LoRA-DLC

A Gradio-based demonstration for the Tongyi-MAI/Z-Image-Turbo diffusion pipeline, enhanced with a curated collection of LoRAs (Low-Rank Adaptations) for style transfer and creative image generation. Users can select from 27 pre-loaded LoRAs (e.g., Turbo Pencil, Ghibli Style, Pixel Art) or add custom ones from Hugging Face repositories. Generates high-quality images from text prompts, with trigger words automatically integrated for optimal results. Supports optimizations like AoTI compilation and FA3 for faster inference.

• LoRA Gallery: Interactive selection from 27 specialized LoRAs, each with preview images, trigger words, and direct links to Hugging Face repos.
• Custom LoRA Support: Input any Hugging Face repo (e.g., "Shakker-Labs/AWPortrait-Z") to dynamically load and use new styles; auto-detects weights, trigger words, and previews.
• Prompt Integration: Automatically prepends or appends LoRA trigger words to prompts for seamless style application.
• Advanced Controls: Adjustable steps (1-50), seed randomization, LoRA scale (0-3), resolution (up to 1536x1536), and CFG scale (forced to 0.0 for Turbo mode).
• Optimizations: Applies AoTI compilation and FA3 for reduced memory and faster generation on CUDA.
• Custom Theme: OrangeRedTheme with gradients, enhanced typography, and responsive CSS for a polished UI.
• Queueing and Progress: Handles up to 60 concurrent jobs with tqdm-tracked progress bars.
• Base Model Fallback: Generates without LoRAs using the pure Z-Image-Turbo pipeline.

• Python 3.10 or higher.
• CUDA-compatible GPU (recommended for bfloat16; falls back to CPU but slower).
• pip >= 23.0.0 (see pre-requirements.txt).
• Stable internet for initial model/LoRA downloads from Hugging Face.

1. Clone the repository:

   git clone https://github.com/PRITHIVSAKTHIUR/Z-Image-Turbo-LoRA-DLC.git
   cd Z-Image-Turbo-LoRA-DLC
   

2. Install pre-requirements (for pip version): Create a pre-requirements.txt file with the following content, then run:

   pip install -r pre-requirements.txt
   

   pre-requirements.txt content:

   pip>=23.0.0
   

3. Install dependencies: Create a requirements.txt file with the following content, then run:

   pip install -r requirements.txt
   

   requirements.txt content:

   git+https://github.com/huggingface/diffusers.git@refs/pull/12790/head
   huggingface_hub
   gradio==6.1.0
   sentencepiece
   transformers
   torchvision
   accelerate
   kernels
   spaces
   torch
   numpy
   peft
   

4. Start the application:

   python app.py
   

   The demo launches at http://localhost:7860 (or the provided URL if using Spaces).

1. Select LoRA: Browse the gallery and click a preview (e.g., "Turbo Pencil") to load it; the prompt placeholder updates with the style name.

2. Enter Prompt: Type a description (e.g., "a serene mountain landscape"); trigger words (e.g., "pencil sketch") are auto-added if applicable.

3. Configure Settings:

   • Expand "Advanced Settings" for steps (default 9), seed, resolution, and LoRA scale (default 0.95).
   • For custom LoRAs, enter a repo path (e.g., "Shakker-Labs/AWPortrait-Z") and press Enter.

4. Generate: Click "Generate" or submit the prompt; monitor the progress bar.

5. Output: View the generated image; download or regenerate with new seeds.

• Select "Ghibli Style" LoRA.
• Prompt: "a whimsical forest adventure".
• Settings: 1024x1024, 9 steps, seed 42.
• Output: Ghibli-inspired image with trigger "Ghibli Style" integrated.

(Full list in code; supports 27 styles like Pixel Art, 80s Horror, etc.)

• LoRA Loading Errors: Ensure repo has .safetensors; check console for warnings. Custom repos must match Z-Image-Turbo base.
• Optimization Fails: AoTI/FA3 requires compatible hardware; fallback to standard pipeline without errors.
• OOM on GPU: Reduce resolution/steps or use low_cpu_mem_usage=True; clear cache with torch.cuda.empty_cache().
• Custom LoRA Invalid: Verify Hugging Face path; must contain model card with instance_prompt or detectable weights/image.
• Generation Slow: Turbo mode uses 0.0 CFG; increase steps for quality but expect longer times.
• UI Issues: CSS targets gallery/buttons; set ssr_mode=True if rendering fails.
• Diffusers Branch: Uses PR #12790; update via git if conflicts.

Contributions welcome! Fork the repo, add new LoRAs to the loras list, or enhance UI/optimizations, then submit PRs with tests. Ideas:

• More LoRA presets.
• Img2Img support (currently ignored for Turbo).
• Batch generation.

Repository: https://github.com/PRITHIVSAKTHIUR/Z-Image-Turbo-LoRA-DLC.git

Apache License 2.0. See LICENSE for details.

Built by Prithiv Sakthi. Report issues via the repository.