LLaDA-Image-Turbo — ComfyUI / RebelAI Quantization
Optimized weights for LLaDA-Image-Turbo intended for the RebelAI ComfyUI integration.
This repository supports both text-to-image generation and native LLaDA image editing.
Repositories
- ComfyUI nodes / GitHub: https://github.com/RealRebelAI/LLaDa-Image_ComfyUI
- Model weights / this Hugging Face repository: https://huggingface.co/realrebelai/LLaDa-Image-Turbo_ComfyUI
- Official base model: https://huggingface.co/inclusionAI/LLaDA-Image-Turbo
- Official LLaDA-Image source: https://github.com/inclusionAI/LLaDA-Image
Base Model / Quantization Tree
This repository is a quantized derivative of:
inclusionAI/LLaDA-Image-Turbo
The model-card metadata intentionally contains:
base_model: inclusionAI/LLaDA-Image-Turbo
base_model_relation: quantized
This associates the repository with the main LLaDA-Image-Turbo model as a quantized derivative in the Hugging Face model relationship/quantization tree.
Available Weights
| File | Format | Description |
|---|---|---|
LLaDA-Image-Turbo-transformer-BF16.safetensors | BF16 Safetensors | Full transformer weights |
LLaDA-Image-Turbo-transformer-INT8.safetensors | INT8 Safetensors | Native INT8 transformer |
LLaDA-Image-Turbo-text_encoder-Q4_K_M-v3.gguf | GGUF Q4_K_M | Quantized LLaDA2-MoE text encoder |
Native INT8 Transformer
The INT8 transformer is stored as native Safetensors. It is not a transformer GGUF.
The custom ComfyUI runtime loads the quantized transformer while preserving the LLaDA-Image transformer architecture.
Q4_K_M Text Encoder
The LLaDA2-MoE text encoder is provided as Q4_K_M GGUF for substantially lower storage/runtime memory requirements than the original full text encoder.
Text-to-Image
LLaDA-Image-Turbo is designed for fast generation.
Recommended starting settings in the ComfyUI integration:
- Steps: 4
- CFG / guidance scale: 1.0
The ComfyUI nodes are available here:
https://github.com/RealRebelAI/LLaDa-Image_ComfyUI
Native Image Editing
The same LLaDA-Image-Turbo model also supports native image editing. No separate editing checkpoint is required.
The RebelAI ComfyUI node uses the model's native:
generation_mode="editing"
The source image is processed through LLaDA's image-conditioning/SigVQ path. This is not conventional img2img implemented with a denoise-strength slider.
Typical ComfyUI graph:
LLaDA Image Loader
|
v
LLaDA Image Edit <---- Load Image
|
v
Save Image
Editing inputs include:
- Source image
- Edit instruction
- Width / height
- Steps
- Guidance scale
- Seed
- Optional negative prompt
Editing dimensions should be divisible by 32.
Example instruction:
Turn the fox into a white arctic fox while preserving the forest composition and realistic photography.
ComfyUI Installation
Get the custom nodes here:
https://github.com/RealRebelAI/LLaDa-Image_ComfyUI
Typical model placement:
ComfyUI/
└── models/
├── diffusion_models/
│ ├── LLaDA-Image-Turbo-transformer-INT8.safetensors
│ └── LLaDA-Image-Turbo-transformer-BF16.safetensors
├── text_encoders/
│ └── LLaDA-Image-Turbo-text_encoder-Q4_K_M-v3.gguf
└── vae/
└── LLaDa_VAE.safetensors
The ComfyUI integration also loads the supporting LLaDA pipeline components required by the official architecture.
Upstream
All model architecture and original model weights originate from inclusionAI's LLaDA-Image project.
Official LLaDA-Image-Turbo:
https://huggingface.co/inclusionAI/LLaDA-Image-Turbo
Official source:
https://github.com/inclusionAI/LLaDA-Image
Please refer to the upstream project for the original model documentation, research information, and applicable licensing terms.
Credits
- inclusionAI — LLaDA-Image / LLaDA-Image-Turbo
- RealRebelAI — ComfyUI integration, INT8 runtime/weights, and GGUF text-encoder integration
License
These files are derivatives of LLaDA-Image-Turbo. The upstream model's applicable license and usage terms continue to apply. The ComfyUI integration code is maintained separately in the GitHub repository linked above.