LLaDA-Image: InclusionAI Open 6B Image Generation and Editing
InclusionAI releases LLaDA-Image, a 6B unified generation and editing model with a 4-step Turbo variant, plus a community ComfyUI node pack with INT8 and GGUF weights.
Photorealistic generation samples from the official repository: natural lighting, lifelike detail and coherent multi-subject scenes.
One checkpoint for generating and editing
LLaDA-Image is a unified diffusion model where both the backbone and the DiT are diffusion models trained in a single framework. Unlike editor add-ons bolted onto a text-to-image model, a single checkpoint handles:
- Text-to-image generation in 50 steps on the Base model
- Instruction-guided editing with reference preservation, using the model's native editing path rather than a denoise-strength img2img workaround
- VQ-conditioned generation through the SigVQ conditioning module
- Chinese and English text rendering in generated images
The training recipe is fully open in the accompanying arXiv report: image-only pre-training builds the visual prior first, paired language supervision comes next, and joint generation-editing training unifies the two skills. On Qwen-Image-Bench it scores 53.53 in English and 53.38 in Chinese, state-of-the-art overall results for an open 6B model at release.
Chinese and English text rendering with poster-grade layouts.
Native editing: content stays faithful while only the instructed change is applied.
Turbo: 4 steps instead of 50
Alongside the Base model, LLaDA-Image-Turbo distills the pipeline with Twin-DMD distillation down to 2 to 4 sampling steps for both generation and editing. Four checkpoint variants ship on day one: BF16 and FP8 for each of Base and Turbo, about 49 GB per variant in diffusers layout.
Qwen-Image-Bench overall scores: LLaDA-Image tops the open 6B class in both English and Chinese.
Run it in ComfyUI with the RebelAI node pack
A community integration from RealRebelAI landed the same day the weights dropped. The pack provides four nodes: LLaDA Image Loader, LLaDA Image Text to Image, LLaDA Image Edit and LLaDA Image Unload. The edit node calls the model's native editing path (generation_mode="editing"), not a denoise-strength approximation, and edit dimensions must be divisible by 32.
The integration uses optimized weights from RealRebelAI/LLaDa-Image-Turbo_ComfyUI and a matching Base pack, about 29 GB per pack:
| File | Placement | Size |
|---|---|---|
LLaDA-Image-Turbo-transformer-BF16.safetensors | models/diffusion_models/ | 13.1 GB |
LLaDA-Image-Turbo-INT8.safetensors | models/diffusion_models/ | 6.6 GB |
LLaDA-Image-Turbo-text_encoder-Q4_K_M.gguf | models/text_encoders/ | 9.2 GB |
LLaDa_VAE.safetensors | models/vae/ | 0.17 GB |
The INT8 transformer is a native safetensors quantization rather than GGUF; GGUF is used only for the quantized LLaDA2-MoE text encoder. For Turbo, start with 4 steps and CFG 1.0.
Training code is listed as coming soon on the official roadmap.
Comments
Sign in with GitHub to join the discussion.