Many image models treat generation, editing, and visual understanding as separate problems; serving all three in a consumer product makes that separation costly. The paper’s central bet is that shared attention can connect instructions, reference images, and target image latents without forcing understanding and generation to use the same parameters. That design aims to preserve specialization while keeping the overall system practical to serve.
WeLM-Image: A Unified Image Generation Model towards Visual Intelligence
Combines multimodal understanding with text-to-image generation, instruction-guided editing, and multi-reference synthesis. Its MoT–MoE design separates understanding and generation parameters while sharing attention; an eight-step Turbo variant targets single-GPU serving.
Introduction
Information
- Websitearxiv.org
- OrganizationsWeChat Vision, Tencent Inc.
- AuthorsCanmiao Fu, Chen Li, Guohui Zhang, Hao Lu, Hongyang Wei, Huadong Wu, Huaisong Zhang, Hubery Yin, Liyuan Cui, Qixin Yan …
- Published date2026/11/05
Categories
More Items
Reduces cumulative visual drift in repeated image edits by applying preservation corrections near the end of the denoising trajectory, while aiming to retain the base model’s single-turn capabilities.
Predicts structured geometry such as bounding boxes and robot trajectories in parallel, rather than generating coordinates one at a time. It reduces errors on evaluated robot-planning tasks while cutting decoding time.
Lets robot manipulation policies draw on visual representations from different depths of a vision-language backbone instead of relying on a separate encoder’s fixed output. The approach improves control results on LIBERO and real-world tasks.