LLaDA-Image is a unified image generation and editing system built by pairing a 6B diffusion transformer trained from scratch with a frozen vision-language understanding module. Instead of starting from paired image-text data, the authors first build a visual generation prior using image-only pre-training and mid-training on a pipeline of 220M samples, most of them real images, then align the model with language and fine-tune it for editing. On Qwen-Image-Bench, it scores 53.53 on English and 53.38 on Chinese, the best open-source results on both tracks. A distilled variant, LLaDA-Image-Turbo, generates in 2 to 4 sampling steps. The authors release model weights, training code, and detailed recipes.
Image-only pretraining helps build a strong open image generator
LLaDA-Image combines a six-billion-parameter diffusion transformer with a frozen vision-language module, while its distilled version generates images in two to four sampling steps.
Independent
Chuyan Chen · Haoxing Chen · Kun Chen · Zhenglin Cheng · Long Cui · Ruishan Fang · +24 more
Research Digest··2 min read
In one line
A 6B diffusion transformer trained from scratch with a mostly real 220M-sample pipeline achieves open-source state-of-the-art Qwen-Image-Bench scores of 53.53 and 53.38.
Before you read this
Two ideas this paper builds onunified multimodal understanding and generationA single end-to-end model processes and generates multiple data types (e.g., images, text) with equal competence.24 papers →vision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.216 papers →
Read it as
The authors trained a 6B diffusion transformer from scratch, first using image-only pretraining and mid-training to establish a visual generative prior before relying more heavily on image-text supervision.
The authors built LLaDA-Image around a 6B Diffusion Transformer, or DiT, paired with a frozen vision-language understanding module based on the LLaDA2.0-Mini diffusion language model.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.
What it released
CodeWeights