WeVisOne-VL unifies object- and pixel-level visual perception
Original title: WeVisOne-VL: Unified Visual Perception from Objects to Pixels
A pixel decoder behind an <OBJ> token emits boxes, masks, and mattes after curriculum SFT and mixed-task RL.
WeChat researchers present WeVisOne-VL, a vision-language model meant to handle fine-grained perception that current VLMs handle unreliably. Structured next-token prediction covers detection, referring localization, counting, OCR, segmentation, and matting, from points and boxes to masks and alpha mattes. A lightweight pixel decoder is tied in through a special <OBJ> token. Training is three-stage curriculum supervised fine-tuning followed by mixed-task reinforcement learning. The authors report leading results on more than 30 benchmarks and say the model is released open source.
- Tasks span detection, localization, counting, OCR, segmentation, and matting.
- The pixel decoder produces instance-level masks and alpha mattes.
- Training is three-stage curriculum SFT, then mixed-task RL.
- Authors report leading scores on more than 30 benchmarks.
- The model is described as released to the open-source community.
- Organizations
- AI models
- WeVisOne-VL
From the original article
1 min read at huggingface.co
Existing vision-language models (VLMs) remain unreliable for fine-grained visual perception. In this report, we present WeVisOne-VL (WeChat Vision All-in-One), a powerful VLM that unifies object-level and pixel-level perception through structured next-token prediction.
By Tianming Liang, Junyi Pan, Yuxi Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.