✦
Summary
WeVisOne-VL is a vision-language model that handles fine-grained visual perception from individual objects down to pixels. A Qwen3-VL backbone predicts structured coordinate tokens, while a lightweight pixel decoder converts object tokens into detailed masks and alpha mattes.
Key Features
Object detection and referring localization.
Object counting and pointing.
OCR for text in images.
Instance segmentation masks.
Continuous alpha matting.
A 9.6M-parameter pixel decoder.
Normalized coordinate vocabulary.
Absent-object rejection.
About WeVisOne-VL
It gives developers a unified perception component for visual interfaces, image analysis, and object-aware editing pipelines. Core capabilities include Object detection and referring localization, Object counting and pointing, OCR for text in images, and Instance segmentation masks.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Continuous alpha matting, A 9.6M-parameter pixel decoder, Normalized coordinate vocabulary, and Absent-object rejection.