LLaDA-Image ships 6B open image model with full training recipe

LLaDA-Image, a 6B diffusion transformer released as an arXiv preprint with weights, training code and full recipe, reports state-of-the-art scores among open-source image generators on Qwen-Image-Bench, hitting 53.53 on the English track and 53.38 on the Chinese track.

The authors pair the DiT with a frozen vision-language module built on the LLaDA2.0-Mini backbone and skip paired image-text data at the outset. "Instead of relying heavily on paired image-text data from the beginning, we first build a strong visual generative prior through image-only pre-training and mid-training," the paper says. The pipeline runs on 220M samples, uses parameter-free RMSNorm throughout the DiT, and trains with the Muon optimizer.

A distilled variant, LLaDA-Image-Turbo, brings inference down to 2-4 sampling steps. The paper argues the base model "produces highly photorealistic images while accurately following fine-grained editing instructions," and the authors say they "release our model weights, training code, and detailed recipes" to support follow-on work.