WeLM-Image: 31.7B unified image model for WeChat generates in 4.98 s
A new paper introduces WeLM-Image, a unified image foundation model that brings multimodal understanding into three tasks: text-to-image generation, instruction-guided editing and multi-reference synthesis. It is designed for single-GPU deployment in WeChat's large-scale consumer applications. The model treats text, reference images and target latents as one interleaved sequence.
The architecture is a Mixture-of-Transformers (MoT) backbone. It combines shared attention with independent understanding and generation parameters, which the authors say limits optimization interference between the two. Each path also carries Mixture-of-Experts (MoE) modules for sparse capacity. In total the model stores 31.7B parameters, and about 5.4B generation parameters are activated per target latent token.
On the data side, the authors describe a generation-oriented data engine. It combines a knowledge-grounded visual vocabulary with large-scale visual clustering to balance semantic coverage and editing tasks, while training-time evaluation guides iterative data refinement. Bilingual annotations and verified editing supervision support progressive training up to native 1.5K resolution.
Post-training works in two steps. Generation and editing are optimized separately with rubric- and checklist-based rewards. The specialized policies are then consolidated through noise-aware on-policy distillation.
For results, the paper says that in the reported overall comparisons the 1536-resolution model ranks first among the evaluated open-source models on PRISM-Bench, Qwen-Image-Bench and UnicBench. The ranking covers only those open-source models and those reported comparisons.
Finally, MeanFlow-based few-step distillation produces WeLM-Image-Turbo. It generates a 1024x1024 image in 4.98 s on a single NVIDIA H20 GPU using eight guidance-free steps.
Key facts
- WeLM-Image is a unified image model for text-to-image generation, instruction-guided editing and multi-reference synthesis, designed for single-GPU deployment in WeChat's consumer applications.
- It uses a Mixture-of-Transformers backbone with Mixture-of-Experts modules: 31.7B stored parameters, about 5.4B generation parameters activated per target latent token.
- Progressive training reaches native 1.5K resolution; the 1536-resolution model ranks first among evaluated open-source models on PRISM-Bench, Qwen-Image-Bench and UnicBench, per the authors' reported comparisons.
- MeanFlow-based few-step distillation yields WeLM-Image-Turbo, which makes a 1024x1024 image in 4.98 s on one NVIDIA H20 using eight guidance-free steps.
Why it matters
Most image systems treat generation and editing as separate jobs. WeLM-Image folds generation, instruction-guided editing and multi-reference synthesis into one model that reads text, reference images and target latents as a single sequence. The stated design target is practical: single-GPU deployment in WeChat's large-scale consumer applications. The sparse design (31.7B stored, about 5.4B generation parameters active per target latent token) and the eight-step Turbo variant both point at serving cost.
Who it affects
The model is designed for WeChat's consumer applications, so the people closest to it are WeChat users and the engineers who serve image features at that scale. Researchers working on unified understanding-and-generation models, sparse mixtures and few-step distillation get a described recipe: MoT with MoE, rubric-based rewards, noise-aware on-policy distillation and MeanFlow distillation.
How to use it
The abstract describes the method and results, not a release. No weights, code or availability are stated in the source, so there is nothing to run yet. For practitioners, the useful part is the recipe: separate reward optimization for generation and editing followed by consolidation, and few-step distillation to eight guidance-free steps for fast inference.
How solid is it
The claims are the authors' own. The first-place ranking is stated for the evaluated open-source models on PRISM-Bench, Qwen-Image-Bench and UnicBench, in the reported overall comparisons. Benchmark scores and the names of the competing models are not given in the source text. The 4.98 s figure belongs to WeLM-Image-Turbo (1024x1024, one H20, eight steps), not to the base model.
Risks and caveats
Ranking first is not claimed against closed-source models. The source does not say the model is already deployed in WeChat, only that it is designed for such deployment. No training compute, dataset size or release terms are given, and the results are self-reported, so independent checks are still to come.