A unified visual perception model from Tencent's WeChat Vision team, built on Qwen3-VL: a single model that performs open-vocabulary detection, referring expression localization, counting, pointing, OCR localization, as well as pixel-level referring segmentation and promptable matting (alpha matting) using text prompts.
Rising score 65 · Likes 67 · Downloads 3786
Reasons
- Boxes and points are generated directly using dedicated coordinate tokens (\u003c0\u003e to \u003c1000\u003e, normalized coordinates), returning None when the target is absent; for dense tasks, an instance token is output after each box, and a pixel decoder with 9.6M parameters generates full-resolution alpha masks.
- Detection, segmentation, and matting share the same predict interface, switched via the task parameter, with the query specifying the object to find.
- There is a paper and a project page; released on 22 10, and not included in previous issues. The rising score of 65 only reflects attention level.
Usage
- After installing the inference package from GitHub Tencent/WeVisOne-VL: model = WeVisOneVL("tencent/WeVisOne-VL-4B", device="auto", dtype="bfloat16"); result = model.predict(image, task="matting", query="person", temperature=0.0), where result.masks contains the masks.
- The mask_decoder.pt file in the repository is a required component and must be placed alongside the main weights.
Risks
- The model card provides no evaluation scores, memory requirements, or a complete list of task names, offering only a matting example.
- The official documentation states that fine edges are sensitive to resolution, severe occlusion, transparent objects, and prompt wording; it supports only single-turn, single-image inputs, with videos rejected by the pixel decoder; it is positioned for research purposes and not for identifying personal identities.
- The pixel decoder is in .pt format (which may contain pickle data); ensure it comes from the official repository before loading.