開源模型
LLaDA-Image Opens Up Bilingual Chinese-English Generation and Editing, With Turbo Cutting Sampling to Four Steps
InclusionAI has released the Base, Turbo, and FP8 weights for LLaDA-Image, enabling a single model to support text-to-image generation, reference-image editing, and Chinese and English text rendering. The paper details a progressive training method using 220 million samples, but the promised training code has yet to be delivered.
On September 4, InclusionAI released the model weights and inference code for the LLaDA-Image family. At its core is a 6B Diffusion Transformer trained from scratch; after the full pipeline adds the frozen LLaDA 2.0 Mini vision-language module, Hugging Face lists the complete checkpoint at approximately 7B parameters. Text prompts are converted into DiT conditioning through a Residual Query Adapter, an understanding model, and a connector. For image editing, the reference image is fed directly into the DiT as SigLIP-VQ features and clean VAE latents, avoiding the loss of composition and pixel-level details that can occur when relying solely on semantic encoding.
The data pipeline is the more technically significant aspect. The team first used image self-conditioning to establish visual priors, then introduced paired image-text data and joint generation-and-editing training. Of roughly 220 million generation samples in total, 98% are real images, with more than 90% used during the image-only modality stage. The model was stabilized during training using RMSNorm with no trainable parameters and the Muon optimizer. The Turbo version was then distilled using TwinFlow/distribution matching, reducing the Base model’s recommended 50 sampling steps to two to four. In the authors’ evaluation on Qwen-Image-Bench, Base achieved overall scores of 53.53 in English and 53.38 in Chinese, which they claim ranks first among open-source models; Turbo trades some quality for lower latency.
For deployment, developers can load the BF16 or FP8 weights directly with PyTorch 2.8 and Diffusers 0.39. A single checkpoint supports text-to-image generation, VQ-conditioned generation, and instruction-based editing. Two caveats remain: the benchmark results have not yet been independently reproduced, and although the repository claims to provide a “fully open training recipe,” it contained only inference code and weights at publication time, with the training code still marked “coming soon.” The next points to watch are whether the training code is sufficient to reproduce every stage and how much FP8 degrades Chinese glyph rendering, long-text rendering, and reference-image consistency in practice.