The model casts all tasks as structured next-token prediction.
A vision-language model unifies object and pixel-level perception with structured token prediction
Tencent's WeVisOne-VL handles detection, referring, OCR, segmentation and matting through one interface, using a special token to trigger a lightweight pixel decoder.
Chinese Tech
Tianming Liang · Junyi Pan · Yuxi Li · Canmiao Fu · Zhipeng Huang · Chong Sun · +2 more
WeChat Vision, Tencent Inc.
Research Digest··2 min read
Tianming Liang, Junyi Pan, Yuxi Li and colleagues at Tencent present WeVisOne-VL, a vision-language model that unifies object-level and pixel-level perception in one interface.
§