What they did

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. Training begins with image-only pretraining and mid-training rather than depending on paired image-text data from the outset.

The DiT uses parameter-free RMSNorm and the Muon optimizer. The authors also distilled the model into LLaDA-Image-Turbo for inference with two to four sampling steps.

Key findings

• LLaDA-Image scored 53.53 on the English track and 53.38 on the Chinese track of Qwen-Image-Bench.
• The authors report that these are the highest overall scores among the open-source models evaluated on both tracks.
• The same framework supports image generation and fine-grained instruction-based editing, with the authors reporting photorealistic outputs and accurate instruction following.
• LLaDA-Image-Turbo reduces generation to two to four sampling steps, although the abstract does not provide corresponding quality or latency measurements.

Why it matters

The work suggests that a strong visual prior can be developed through image-only training before extensive image-text alignment, potentially reducing dependence on paired datasets. Releasing weights, code, and detailed recipes also makes the training and optimization choices more inspectable and reproducible than those of closed image generators.

Caveats

The supplied abstract does not describe benchmark baselines, human evaluations, compute requirements, data provenance, or safety testing. It states that the generation pipeline contains 220 million samples but says only that “98” are real images, without a unit or percentage sign, leaving the dataset composition unclear; the claimed open-source state of the art therefore requires confirmation from the full experimental details.