WeVisOne-VL Unifies Object and Pixel-Level Visual Perception
October 21, 2026
WeVisOne-VL integrates a lightweight pixel decoder into a VLM using a special <OBJ> token to enable unified object detection, segmentation, and matting. The model uses a three-stage curriculum SFT and mixed-task RL to achieve leading performance across 30 benchmarks.
HOW THIS AFFECTS YOU
●
builderYou can use a single model for tasks ranging from bounding boxes to alpha mattes.
●
researcherThe use of <OBJ> tokens to bridge object-level and pixel-level tasks provides a new architectural pattern for unified VLMs.