Original title: LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes
Chuyan Chen, Haoxing Chen, Kun Chen, Zhenglin Cheng +26 · September 2026
Creating high-quality image generators often relies on massive sets of paired pictures and captions. To improve this process, researchers created LLaDA-Image, a unified image generation system. Instead of depending heavily on image-text pairs right away, the team first taught the model visual patterns using only images during early training. The full setup connects a new image-making model with a frozen language system that interprets text. They also distilled this model into a faster version called LLaDA-Image-Turbo, which makes pictures in just 2-4 steps. Overall, LLaDA-Image creates highly photorealistic pictures, accurately follows detailed editing prompts, and reached a new state-of-the-art score among open-source models on both the English and Chinese tracks of an image evaluation benchmark.
Imagine you're learning to paint by copying famous artworks on your own to get a feel for colors and shapes. Then, a friend says, 'Add more light here,' but applying that without ruining the entire painting requires understanding how the whole composition balances. In this paper, the system first learns from images alone to master visual patterns, then connects to a language model to follow editing commands, allowing it to adjust details precisely without starting over.
Open-source image generators often trail behind proprietary ones, but this model matches top performance while being fully open. This allows anyone to access powerful tools for creating and editing images without proprietary barriers.
The abstract does not explicitly state limitations, but the work is focused on open-source models and evaluated on specific benchmarks like Qwen-Image-Bench, which may not cover all real-world scenarios. It also relies on particular training data and methods, so generalizability is not discussed.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes — Chuyan Chen, Haoxing Chen, Kun Chen et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.