生成式影像模型
LLaDA-Image Unifies Image Generation and Editing in a Single 6B Diffusion Model, With Turbo Cutting Sampling to 4 Steps
Ant Group’s inclusionAI has released LLaDA-Image weights and Diffusers inference code. A single model handles text-to-image generation, reference-image editing, and Chinese and English text rendering. The distilled version uses Twin-DMD to reduce the recommended sampling process from 50 steps to 4, but training code and independent performance validation are still unavailable.
On September 4, inclusionAI released the weights and inference code for LLaDA-Image Base, Turbo, and FP8. The paper describes it as a unified 6B-parameter image model: its generation backbone is a Diffusion Transformer trained from scratch, while prompt understanding is handled by a frozen LLaDA2.0-Mini vision-language module. The same checkpoint can perform text-to-image generation, VQ-conditioned generation using discrete visual tokens produced by LLaDA2, and instruction-based editing that preserves content from a reference image, without requiring a separate, specialized editing backbone.
Rather than relying entirely on paired image-text data from the outset, the training pipeline first establishes visual priors through image-only pretraining and mid-training, then introduces language supervision and joint generation-and-editing training. The authors say the full pipeline uses 220 million samples, including 98 million real images. The DiT uses parameter-free RMSNorm and the Muon optimizer. The Turbo version is distilled from Base using Twin-DMD, shortening the 50-step process to a recommended 4 steps. Official BF16 and FP8 versions are also available to reduce the memory required to load the weights.
The authors report scores of 53.53 for English and 53.38 for Chinese on Qwen-Image-Bench, claiming the best results among open-source models. This could appeal to teams that need Chinese text layout and a shared pipeline for generation and editing, but the figures currently come primarily from the authors’ own evaluations. Latency at different resolutions, peak VRAM usage, and the quality degradation associated with FP8 have not yet been disclosed. More importantly, although the paper says the complete training code has been released, the project page still labels it “coming soon.” Hugging Face metadata also indicates approximately 7B parameters, which does not fully align with the paper’s 6B description. Engineering teams should wait for the training code, a detailed explanation of the model’s components, and third-party tests of Chinese and English text rendering and editing consistency before assessing its reproducibility.