Tencent WeLM-Image unifies text-to-image, editing at 31.7B

Tencent's WeChat Vision group has published WeLM-Image, a 31.7B-parameter image model that activates roughly 5.4B generation parameters per target latent token and, in a distilled Turbo variant, renders a 1024x1024 image in 4.98 seconds on a single NVIDIA H20 GPU using eight guidance-free steps.

The paper, WeLM-Image: A Unified Image Generation Model towards Visual Intelligence, puts text-to-image, instruction-guided editing, and multi-reference synthesis into one backbone. "A Mixture-of-Transformers (MoT) backbone combines shared attention with independent understanding and generation parameters to limit optimization interference," the authors write. Each path carries its own sparse Mixture-of-Experts capacity, and the editing side covers 27 action types spanning object, background, portrait and restoration work.

On benchmarks the authors report the 1536-resolution model "ranks first among the evaluated open-source models on PRISM-Bench, Qwen-Image-Bench, and UnicBench." The abstract names no closed-source comparators and publishes no per-benchmark scores.

The design target is spelled out: "single-GPU deployment in WeChat's large-scale consumer applications." WeLM-Image lands into a steady run of Chinese generative releases we have been tracking, with our open-source feed logging 355 stories in the last 90 days.