LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes

LLaDA-Image. Building Strong Image Generators with Fully Open Training Recipes. By Chuyan Chen and twenty-nine others, from the AGI Research Center at Inclusion AI. Published on the third of September. Its code repository is public but slight: seven source files.

What now exists is an open image generator released with the parts that usually stay private. Model weights, training code, and the detailed recipes. The framework pairs a six billion parameter diffusion transformer, trained from scratch, with a frozen vision-language understanding module built on a diffusion language model backbone.

The training strategy is the substantive claim. Rather than leaning on paired image-text data from the beginning, the authors first build a strong visual generative prior through image-only pre-training and mid-training, and bring the text pairing in afterwards. Two efficiency choices go alongside it: parameter-free RMSNorm used throughout the transformer, and the Muon optimizer. The pipeline runs to two hundred and twenty million samples.

The composition of that data is stated, which is the sort of detail an open recipe exists to provide and most releases omit. The distilled variant matters too: the model is distilled into a turbo version that runs inference in two to four sampling steps, so the release covers both the quality configuration and the fast one rather than leaving the speedup as an exercise.

On Qwen-Image-Bench it reaches overall scores of fifty-three point five three on the English track and fifty-three point three eight on the Chinese track, setting a new state of the art among open-source models on both. Being evaluated on both tracks matters for who this is for: a generator strong only in English is a narrower resource than the headline suggests.

The contribution is as much the disclosure as the model. Image-only pre-training before text pairing is a recipe claim rather than an architecture claim, and it is only checkable because the code and the recipe came with the weights.

LLaDA-Image. Building Strong Image Generators with Fully Open Training Recipes. By Chuyan Chen and twenty-nine others, from the AGI Research Center at Inclusion AI. Published on the third of September. This is the long version: the whole paper rather than the abstract.

This is a report rather than a result, and it says so. The question it sets is a practical systems one: what does it take to train a strong image generator from scratch with a fully open recipe? Everything else follows from taking that question literally, which is rarer than it sounds.

The scope statement is unusually disciplined and deserves quoting in substance. The authors trace the complete path from image-only visual-prior learning through language alignment and joint generation and editing, to distillation. For each stage they document the data construction, the model interfaces, the optimization choices, the training schedule, and the evaluation used to produce the released checkpoints. The stated goal is to make the design trade-offs inspectable and the recipe reusable. And then the disclaimer that most papers omit: they do not claim every component is universally optimal, nor do they attempt an exhaustive comparison of all architectures or data mixtures.

The architecture pairs a six billion parameter diffusion transformer, trained from scratch, with a frozen vision-language understanding module built on a diffusion language model backbone. The training strategy is the substantive claim. Rather than relying heavily on paired image-text data from the start, they first build a strong visual generative prior through image-only pre-training and mid-training, and bring language alignment in afterwards. The generation pipeline runs to two hundred and twenty million samples. Two efficiency choices sit alongside: parameter-free RMSNorm used throughout the transformer, and the Muon optimizer, both chosen for efficient and scalable optimization rather than for quality.

The distilled variant matters as much as the main model for a release like this. The model is distilled into a turbo version running inference in two to four sampling steps, so what is published covers both the quality configuration and the fast one. A release that documented only the expensive path would leave the most practically relevant question, how to make it cheap, as an exercise.

On Qwen-Image-Bench the model reaches overall scores of fifty-three point five three on the English track and fifty-three point three eight on the Chinese track, setting a new state of the art among open-source models on both. Evaluating on both tracks is doing real work rather than adding a number: a generator strong only in English is a much narrower resource than the headline would suggest, and bilingual text rendering is among the harder things for an image model to get right. Model weights, training code and detailed recipes are released, which for a report whose stated contribution is reusability is the part that determines whether the rest of it counts for anything at all.

There is no limitations section, which is a gap, though the scope statement does much of the same work by declining to claim optimality. What can be established from the text is narrower than the state-of-the-art framing implies. The comparison is against open-source models on one benchmark family, and the image-only-first training claim is a property of this recipe rather than something ablated against the alternative. The reader challenge is explicitly not evidence. And the deepest limitation of any recipe report is that it documents one path through a space it never mapped, which the authors say outright rather than leaving to be inferred.

The contribution is the disclosure as much as the model. Image-only pre-training before language alignment is a recipe claim rather than an architecture claim, and a recipe claim is only checkable because the code, the schedules and the data construction came with the weights. That is a different kind of contribution from a better number, and it is one that only counts if the documentation is honest about what it does not establish. On that count this one is unusually careful, and the care is what makes the recipe worth reading rather than merely available.