Tencent/WeVisOne-VL

What is WeVisOne-VL?

WeVisOne-VL (short for WeChat Vision All-in-One) is a vision-language model from Tencent that handles many different visual tasks with a single system. Instead of needing separate tools for each job, one model can detect objects, count them, read text, segment images, and more.

It works at two levels of detail:

• Object-level: finding objects with boxes or points, counting them, reading text (OCR), and locating objects described in natural language.
• Pixel-level: producing binary masks (which pixels belong to an object) and continuous alpha mattes (soft edges, e.g., for hair or fur), useful for cutouts and image editing.

How it works

The model has three main parts:

1. Data Engine: A knowledge-graph-driven system that automatically annotated over 100 million images. It discovers concepts, localizes them with specialized tools verified by a vision-language model, and extends the results to masks and alpha mattes.

2. Unified Task Formulation: Every task is expressed as structured text. Coordinates use tokens <0> to <1000> for boxes and points, None marks absent targets, and an <OBJ> token signals pixel-level output.

3. Model Architecture: A lightweight pixel decoder (only 9.6M parameters) combines the hidden state of each <OBJ> token, its predicted box, and multi-level features from the Qwen3-VL vision transformer to predict a full-resolution mask or alpha matte. No SAM (Segment Anything Model) is required.

Key strengths

• All-in-one: detection, referring, counting, pointing, OCR, segmentation, and matting in a single model.
• Lightweight pixel decoder: 9.6M parameters, no external segmentation model needed.
• Reliable rejection: returns None for absent targets instead of hallucinating objects that aren't there.

Available models and demos

Two model sizes are released on Hugging Face: WeVisOne-VL-2B and WeVisOne-VL-4B. Both have hosted demos you can try, and a Gradio demo can be launched locally.

Supported tasks

Task	What it does	Output
detection	Finds all instances of given categories	Boxes
referring	Locates objects described by a natural-language expression	Boxes or None
point	Marks each queried object with a point	Points
count_box / count_point	Counts matching instances	Boxes/points + count
ocr / ocr_grounding	Finds and recognizes text	Text + boxes
segmentation	Produces binary masks for described objects	Boxes + masks
matting	Produces alpha mattes (soft-edged cutouts)	Boxes + alpha mattes

Performance highlights

On benchmarks, WeVisOne-VL-4B outperforms the strongest open baselines (Qwen3-VL/Qwen3.5 up to 9B, Rex-Omni, LocateAnything, and task-specific models) across all headline metrics, including:

• Detection: F1 Mean 55.3 on LVIS (baseline 50.7), 71.3 on Dense200 (baseline 58.7)
• Referring: F1 Mean 81.8 on RefCOCOg test (baseline 77.6)
• Rejection: 96.4% rejection rate on WeRef-Bench (baseline 73.0)
• Counting: 89.6% accuracy on PixMo-Count test (baseline 82.8)
• OCR: F1 Mean 52.0 on TotalText (baseline 43.3)
• Referring segmentation: gIoU 71.2 on ReasonSeg val (baseline 65.0)
• Box matting: MAD 7.3 on P3M-500-NP (baseline 15.4)
• Text matting: MAD 9.8 on WeRefMat (baseline 22.4)

What's released

• Model weights (2B & 4B) on Hugging Face
• Inference package and Cookbooks (Jupyter tutorials for each task pipeline)
• Gradio web demo
• Benchmark evaluation pipeline
• A matting annotator tool (Mask2Alpha) used to create training data

Not yet released: fine-tuning code (listed as a TODO).

License

Tencent-authored code is under Apache License 2.0. Third-party components keep their original licenses.

Citation

If you use WeVisOne-VL, the README asks you to cite the arXiv paper: WeVisOne-VL: Unified Visual Perception from Objects to Pixels (arXiv:2610.31329).