LLaDA-Image Turbo: 4-Step Generation and Editing by InclusionAI
LLaDA-Image Turbo distills InclusionAI's 6B unified image model to 2-4 steps with Twin-DMD, running text-to-image and native instruction editing in ComfyUI.
LLaDA-Image Turbo
Text-to-ImageImage EditingFew-StepDistilledTwin-DMD distilled variant of InclusionAI's 6B unified image model. Compresses both generation and editing to 2 to 4 sampling steps while keeping the native editing path, and ships as BF16 and INT8 transformer weights for ComfyUI.
| Developer | InclusionAI (Ant Group) |
| Release Date | 2026-09-04 |
| Architecture | Unified diffusion backbone + DiT |
| Parameters | 6B |
| License | Apache-2.0 |
| Inference | 2-4 steps, CFG 1.0 |
Capabilities
- Text-to-image generation: 50 steps on the Base model, 2 to 4 steps on Turbo
- Instruction-guided editing: uses the model's native editing path with reference preservation, not a denoise-strength img2img approximation. Edit dimensions must be divisible by 32
- Bilingual text rendering: legible Chinese and English text in generated images
- VQ-conditioned generation: SigVQ conditioning module
ComfyUI Usage
The community RebelAI node pack provides four nodes: LLaDA Image Loader, LLaDA Image Text to Image, LLaDA Image Edit, and LLaDA Image Unload.
Place the optimized weights in their ComfyUI folders:
| File | Destination |
|---|---|
LLaDA-Image-Turbo-transformer-BF16.safetensors or LLaDA-Image-Turbo-INT8.safetensors | ComfyUI/models/diffusion_models/ |
LLaDA-Image-Turbo-text_encoder-Q4_K_M.gguf | ComfyUI/models/text_encoders/ |
LLaDa_VAE.safetensors | ComfyUI/models/vae/ |
The INT8 transformer is a native safetensors quantization; GGUF is used only for the quantized LLaDA2-MoE text encoder. Start with 4 steps and CFG 1.0.
Related
Guides and workflows related to this model series.
Comments
Sign in with GitHub to join the discussion.