Tencent WeVisOne-VL unifies six vision tasks via <OBJ> token

A single `<OBJ>` token routes six different visual tasks through the same Tencent backbone.

Tencent's WeChat Vision All-in-One team calls the system WeVisOne-VL, 'a powerful VLM that unifies object-level and pixel-level perception through structured next-token prediction.' The six tasks the abstract lists are object detection, referring localization, counting, OCR, segmentation, and matting, with outputs ranging from points and boxes to masks and alpha mattes.

'We integrate a lightweight pixel decoder into the VLM via a special <OBJ> token to generate instance-level pixel outputs,' the authors write. Training is 'a three-stage curriculum SFT followed by mixed-task RL.'

The results claim is single-sourced to the paper itself. 'Across over 30 benchmarks, WeVisOne-VL demonstrates leading performance on both object-level and pixel-level tasks, while maintaining robust rejection for object-absent queries,' the abstract states. No per-benchmark figures and no baseline model names appear in the published text.

The release lands amid a run of open visual work our tracker has been logging this week, next to recent open-source and multimodal releases from Peking University and HKUST.