Tencent releases WeVisOne-VL, unifying object and pixel-level visual perception
A source-linked abstract prepared once when this article was collected.
WeVisOne-VL is a vision-language model that unifies object-level and pixel-level perception through structured next-token prediction. It supports detection, referring localization, counting, OCR, segmentation, and matting, with localization from points and boxes to masks and alpha mattes. It integrates a lightweight pixel decoder via a special <OBJ> token and is trained with a three-stage curriculum SFT followed by mixed-task RL. The authors report leading performance across over 30 benchmarks and robust rejection of object-absent queries, and release the model to the open-source community.
Key takeaways
- Unifies object-level and pixel-level perception via structured next-token prediction in one VLM.
- A lightweight pixel decoder, triggered by a special <OBJ> token, generates instance-level pixel outputs.
- Trained with three-stage curriculum SFT followed by mixed-task RL.
- Reports leading results on over 30 benchmarks and robust rejection of object-absent queries.
Why it matters
Fine-grained, pixel-accurate localization is a known weakness of general VLMs. A single model covering detection through matting, if the reported results hold, could simplify perception pipelines and is released openly for further testing.
Read with care
Source claims not independently checked; benchmark results are self-reported by the authors.
Discussion
Comments are reviewed before publication. Discuss the evidence, not the person.
No published comments yet.
Reader discussion opens with accounts. You can read the original source and published comments now.