LLaDA-Image: An Open Recipe for Unified Image Generation and Editing

Introduction

Open image generators are increasingly being judged not only by visual quality, but also by how well they understand instructions and preserve consistency during editing. The paper LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes presents a unified model family designed for image generation, reference-based editing, and multilingual text rendering.

How the system is built

The core generator is a 6B-parameter Diffusion Transformer trained from scratch. Rather than depending on paired image-text data from the earliest stage, the authors first perform image-only pre-training and mid-training. This image-first route is intended to establish a stronger visual generation prior before language conditions are introduced. The reported training pipeline contains about 220 million samples, including real-image data.

For understanding and conditioning, LLaDA-Image uses a frozen vision-language module built on the LLaDA2.0-Mini diffusion language-model backbone. This creates a division of labor: the DiT models the image generation process, while the frozen module contributes language and multimodal instruction understanding. The resulting family supports text-to-image generation, VQ-conditioned generation, reference-image editing, and Chinese–English text rendering.

The optimization recipe is another notable part of the release. The DiT uses parameter-free RMSNorm throughout and is trained with the Muon optimizer. These choices are presented as a way to make optimization more efficient and scalable. The project also releases model weights, training code, and detailed recipes, giving researchers more than a ready-to-use checkpoint and making further reproduction or modification easier.

Two model variants

• LLaDA-Image uses 50 sampling steps and targets high-quality generation and instruction-guided editing.
• LLaDA-Image-Turbo is distilled for fast inference and editing in 2–4 sampling steps.

Results and significance

The paper reports overall Qwen-Image-Bench scores of 53.53 on the English track and 53.38 on the Chinese track, describing these results as state of the art among open-source models on both tracks. The broader contribution is the attempt to place generation, editing, reference-image control, and bilingual text rendering in one open model family instead of treating them as isolated features.

The available material does not detail training cost, data licensing, or per-capability comparisons. Those factors will matter when the community evaluates reproducibility and real-world robustness. Still, the release is important because it exposes not only model weights but also the implementation and training recipe, helping shift open image-model development from checkpoint sharing toward more reproducible systems research.

Source: Hugging Face Daily Papers