LLaDA-Image

"Building Strong Image Generators with Fully Open Training Recipes." A unified image generator and editor that pairs a 6B Diffusion Transformer trained from scratch with a frozen understanding module on Ant's LLaDA2.0-Mini diffusion-LM backbone. Rather than leaning on paired image-text data from the start, the recipe first builds a strong generator from image-only pre- and mid-training over a 220M-sample pipeline (98M real images), then adds paired supervision; RMSNorm and the Muon optimizer are used throughout, and Twin-DMD distillation yields a 2-to-4-step Turbo model. Ant reports 53.53 EN and 53.38 ZH on Qwen-Image-Bench, an open-source state of the art by its account. Base and Turbo checkpoints plus FP8 variants and inference code shipped September 4, 2026 under Apache 2.0, with training code promised; the paper is from Inclusion AI's AGI Research Center.