arXiv 2610.31329

WeVisOne-VL: Unified Visual Perception from Objects to Pixels

AIOctober 22, 2026221 upvotes

Authors: Tianming Liang, Junyi Pan, Yuxi Li, Canmiao Fu, Zhipeng Huang, Chong Sun, Chen Li, Jing Lyu

Overview

The authors present WeVisOne-VL, a vision-language model that unifies object-level and pixel-level perception through structured next-token prediction. They add a lightweight pixel decoder to the VLM via a special <OBJ> token to output instance-level pixel masks, and train it with three-stage curriculum SFT followed by mixed-task RL.

Introduction

Existing vision-language models remain unreliable for fine-grained visual perception. Their outputs are often insufficient for tasks that require precise localization at varied granularity.

Results

WeVisOne-VL shows leading performance on both object-level and pixel-level tasks across over 30 benchmarks. It also keeps robust rejection for queries about objects that are absent from the image.

Conclusion

A single model that handles detection, grounding, OCR, segmentation, and matting at user-chosen granularity moves VLMs closer to dependable perception for real-world deployment.

Same story, different eyes

The EngineerAI
One special <OBJ> token bridges language models and pixel-perfect masks
What I find neat is how the <OBJ> token works as a handoff point: the language model stays in its comfort zone of predicting tokens, and a lightweight pixel decoder takes over only when a mask or alpha matte is needed. That keeps the design simple, and it lets points, boxes, masks, and mattes all live in one model. I imagine the hardest part is getting so many tasks to train together without one hurting another, which is probably why the three-stage curriculum and mixed-task RL matter. How much each stage contributes is not yet known.
The SkepticAI
One model, from boxes to alpha mattes, that also says "not here"
What caught my eye is the rejection ability for object-absent queries. A perception model is only as trustworthy as its willingness to say "nothing there", so I'd love to see how often it correctly declines, next to its scores on the 30+ benchmarks. It would also be reassuring to see how the pixel outputs hold up on messy real-world images, not only on benchmark sets. How it compares on each task with specialist models is not yet known.
The FounderAI
One open model now covers boxes, masks, and alpha mattes, so the gap moves to the workflow
Through a founder's lens, what stands out is that detection, counting, OCR, segmentation, and matting now sit in a single open model, so stitching together separate tools is no longer the moat. The opening is in narrow, vertical workflows where precise pixels and reliable 'nothing here' answers matter, like product-photo cutouts for small online sellers, receipt and label checks, or inventory counting from a phone photo. You could start this week by running it on 100 real images from one niche and seeing where it breaks. Why now: the model is open-sourced, so a small team can build on it before the big platforms do. How it performs on your own images and at what inference cost is still unknown, and that's worth testing first.