Paper review · October 24, 2026
WeVisOne-VL: A VLM that unifies visual perception from the object level to the pixel level
WeVisOne-VL: Unified Visual Perception from Objects to Pixels
It unifies detection, referring grounding, OCR, segmentation and matting as structured next-token prediction in a single VLM, so the model can output locations at whatever granularity is needed, from boxes to alpha mattes.
Problem
Existing vision-language models (VLMs) are not reliable at fine-grained visual perception. The problems are hallucination of objects that are not there, misreading intent when given unfamiliar expressions, and lower recall from missing less conspicuous objects. Segmentation also often depends on a separate large segmentation model.
Method
The authors built 2B and 4B models on top of Qwen3-VL. Box and point coordinates are output with dedicated tokens for normalized coordinates (<0>~<1000>), and the hidden state of an <OBJ> token generated after each box is fed into a lightweight pixel decoder to produce masks and alpha mattes. The data engine organizes concepts and referring expressions with a knowledge graph; SAM3, an in-house open-vocabulary detector and Rex-Omni propose candidates, and Qwen3-VL verifies them. Masks are produced with SAM2 and alpha labels with a dedicated matting annotator. Model errors are reviewed to regenerate hard negatives and positives. Training uses a three-stage curriculum SFT (task alignment → high-resolution and dense perception → pixel-level tasks), followed by GRPO reinforcement learning on a mix of tasks using the samples the SFT model is most uncertain about.
Results
- The authors report leading performance on both object-level and pixel-level tasks across more than 30 benchmarks (the provided text does not include per-benchmark numbers).
- They report robust rejection of queries whose target is absent, and high recall even in dense scenes.
- The training data inventory contains more than 100 million images and about 3.6 billion bounding boxes.
- A single model supports point, box, mask and alpha matte outputs without a separate large segmentation model.
Limitations
The provided text cuts off before the experiments section, so per-benchmark numbers and baseline comparisons cannot be checked. Most training labels are generated automatically by existing models such as SAM3, SAM2 and Qwen3-VL, with humans spot-checking only a sample, so label quality may be affected by the limitations of those models. Only two model sizes were tested: 2B and 4B.
Why it matters
Pipelines that run separate dedicated models for detection, segmentation and matting could be reduced to a single small VLM. The weights and code are released, so it can be tried in practice right away. The data engine design, which feeds error reviews into the next round of data collection and label correction, is also worth studying.
Key terms
- Referring Localization
- A task of locating the specific object that a natural-language expression refers to, such as "the person holding a red cup on the left."
- Matting
- A task that estimates foreground opacity as a continuous value (alpha) between 0 and 1. It can separate regions with fine boundaries, such as hair, more precisely than a 0/1 mask.
- <OBJ> token
- A special token the model generates after outputting a box. The token's hidden state serves as the representation of that object, which the pixel decoder uses to produce a mask.
- GRPO
- Group Relative Policy Optimization. A reinforcement learning method that generates several answers for the same input and updates the policy by comparing their rewards relative to one another within the group.
Original
Authors: Tianming Liang, Junyi Pan, Yuxi Li, Canmiao Fu, Zhipeng Huang, Chong Sun, Chen Li, Jing Lyu · Posted 2026-10-22 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.