WeVisOne-VL: Unified Visual Perception from Objects to Pixels

一句话要点

腾讯微信团队的视觉语言模型 WeVisOne-VL 用结构化的下一 token 预测统一目标级和像素级感知：检测、指代定位、计数、OCR、分割和抠图都能做，通过 <OBJ> 特殊 token 接一个轻量像素解码器输出掩码；三阶段课程式 SFT 加混合任务 RL，30 多个基准领先，并会拒答图中不存在的目标，模型开源。

Tencent WeChat's WeVisOne-VL unifies object- and pixel-level perception via structured next-token prediction — detection, referring, counting, OCR, segmentation and matting — using an <OBJ> token that drives a light pixel decoder; trained with 3-stage curriculum SFT plus mixed-task RL, it leads 30+ benchmarks, rejects absent-object queries, and is open-sourced.