WeVisOne-VL: Unified Visual Perception from Objects to Pixels

Overview of WeVisOne-VL's unified object- and pixel-level perception capabilities

One model, every granularity: from points and boxes to binary masks and continuous alpha mattes — and a confident None when the target isn't there.

Abstract

Fine-grained perception, reliably.

Existing vision-language models remain unreliable for fine-grained visual perception. We present WeVisOne-VL (WeChat Vision all-in-One), a VLM that unifies object-level and pixel-level perception through structured next-token prediction. It supports object detection, referring localization, counting, pointing, OCR, segmentation and matting — localizing at any user-specified granularity, from points and boxes to masks and alpha mattes.

A lightweight 9.6M-parameter pixel decoder is plugged into the VLM through a special <OBJ> token to produce instance-level pixel outputs. Trained with a three-stage curriculum SFT followed by mixed-task RL on ~100M images, WeVisOne-VL leads 30+ benchmarks across object- and pixel-level tasks while robustly rejecting queries for absent objects.

Interactive

Watch tokens become geometry.

Real outputs from WeVisOne-VL-4B, replayed token by token. Every coordinate is a single vocabulary token <0>…<1000>; every <OBJ> wakes up the pixel decoder.

wevisone_vl · model.predict()

task

prompt

response

ready

Method

A VLM that speaks coordinates — and paints pixels.

Built on Qwen3-VL. No SAM, no external segmentation backbone — the pixel decoder reuses the VLM's own multi-level ViT features.

Click to enlarge

01

Visual Feature Adapter

Taps four levels of pre-merge Qwen3-VL ViT features, projects and refines them with depthwise-conv blocks into a shared pixel feature pyramid.

02

Box-guided Query Decoder

Each <OBJ> hidden state + a Fourier-encoded box prior becomes an instance query, refined by 6 deformable cross-attention blocks.

03

Instance-conditioned Dense Head

Dynamic kernels predict masks, then a coarse-to-fine cascade (S/4 → S/2 → S/1) recovers hair-level, full-resolution alpha mattes.

Coordinates are just tokens.

1,001 dedicated tokens <0>…<1000> encode normalized positions. A box costs exactly four tokens; a point costs two. Absent target? The model simply writes None.

dog

05001000

Data & Training

~100M images. A self-improving data engine.

A knowledge graph of categories and referring expressions drives targeted acquisition; complementary localization experts propose, a VLM verifies, SAM2 extends to masks, and a dedicated annotator produces continuous alpha. Model errors feed back into the next round.

Click to enlarge

SFT · Stage 1

Task alignment

Follow instructions, emit well-formed structured outputs.

90.2M samples

SFT · Stage 2

High-res & dense

Boost recall of small, crowded objects.

40.5M samples

SFT · Stage 3

Pixel-level

Warm-up the decoder, then extend to masks & mattes.

41.8M samples

RL · GRPO

Mixed-task RL

Uncertainty-selected samples; sharper boxes, fewer hallucinations.

178K samples

Results

Leading 30+ benchmarks, from objects to pixels.

Compared with Qwen3-VL / Qwen3.5, Rex-Omni, LocateAnything, specialist detectors, OCR systems and segmentation / matting models.

WeVisOne-VL (ours) Grounding VLMs General VLMs Specialists

Citation

BibTeX

bibtex

@article{liang2026wevisone,
  title   = {WeVisOne-VL: Unified Visual Perception from Objects to Pixels},
  author  = {Liang, Tianming and Pan, Junyi and Li, Yuxi and Fu, Canmiao and
             Huang, Zhipeng and Sun, Chong and Li, Chen and Lyu, Jing},
  journal = {arXiv preprint arXiv:2610.31329},
  year    = {2026}
}