core
LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes
· inclusionAI · 2026-09 · arXiv:2609.03796
Why it matters
A 6B diffusion transformer trained from scratch alongside a frozen vision-language understanding module, released with weights, code and the full training recipe over 220 million samples — an unusually complete open release. The Turbo variant samples in two to four steps via distillation from a 50-step base. By community uptake it is the highest-traction image-generation release of the summer.
What this paper does
read: full textBefore this
Open text-to-image systems coupled visual-prior learning to language alignment from the very start, which meant paired image-caption data was needed during the most compute-intensive stage of training.
The problem
Captions are expensive and lossy. At the low resolutions used early in training, details a caption mentions may simply disappear once the image is downsampled, so the supervision and the training target disagree.
The idea
Decouple the two. Learn the visual prior from images alone, using a frozen vision-language model to extract a semantic condition from the very same crop being trained on, and only bring in paired captions later for alignment.
How it works
A 6B single-stream diffusion transformer is trained from scratch and processes image and condition tokens jointly through shared layers. The understanding side is a frozen LLaDA 2.0-Mini diffusion language model with a SigLIP-VQ vision encoder. A Residual Query Adapter of learnable query tokens cross-attends to the full multimodal sequence, and a shallow transformer connector projects the result into the diffusion transformer's condition space. Training runs image-only at 256 pixels, then image-only at 512, then text-to-image alignment with paired data, then joint generation and editing. More than 90 percent of the generation pipeline is image-only. For editing, reference images bypass the language model entirely and enter through a separate branch carrying SigLIP-VQ features and the clean VAE latent. TwinFlow distillation then compresses sampling to two to four steps by running distribution-matching distillation inside a single shared backbone, using signed time to switch a block between generator and fake-score roles rather than training a second network.
Evidence
Overall scores of 53.53 on the English track and 53.38 on the Chinese track of Qwen-Image-Bench, reported as state of the art among open-source models. Weights, code and the full recipe are released, which is unusually complete.
Limitations
The authors are direct about the weak spots. Counting scores 0.53 on GenEval against 0.73 to 0.91 for baselines. Text rendering trails dedicated models. In editing, perceptual quality outruns semantic consistency, 8.043 against 7.182 on the English track. Layout stability degrades as the number of regions and the length of requested text grow, and 2K generation is an extension rather than an evaluated capability.
Why it matters
It is the strongest evidence so far that the expensive part of training a generator does not need captions at all. If a frozen understanding model can supply the semantic condition from the image itself, the caption bottleneck that shapes most open training pipelines becomes optional.
Abstract, in the authors' own words
We introduce LLaDA-Image, a unified framework that pairs a 6B Diffusion Transformer (DiT) trained from scratch with a frozen vision-language understanding module built on the LLaDA2.0-Mini diffusion language model backbone. 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 generation pipeline comprises 220M samples, 98 of which are real images. For efficient and scalable optimization, we use parameter-free RMSNorm throughout the DiT together with the Muon optimizer. The resulting unified model produces highly photorealistic images while accurately following fine-grained editing instructions. We further distill LLaDA-Image into LLaDA-Image-Turbo, enabling fast inference in 2-4 sampling steps. On Qwen-Image-Bench, LLaDA-Image achieves overall scores of 53.53 and 53.38 on the English and Chinese tracks, respectively, setting a new state-of-the-art among open-source models on both tracks. To support further research on capable and efficient generative models, we release our model weights, training code, and detailed recipes.
Research line
Method note — the shared flow-matching interpolation
Every flow-matching / rectified-flow paper in this atlas trains toward a straight-line path between a noise sample x₀ and a data sample x₁:
Builds on
Nothing recorded yet.
Built on by
Nothing recorded yet.