WeVisOne-VL Connects Bounding Boxes to Pixel Masks with a Compact Decoder

This is a Plain English Papers summary of a research paper called WeVisOne-VL Connects Bounding Boxes to Pixel Masks with a Compact Decoder. If you like these kinds of analyses, you can find more AI and machine-learning research on AIModels.fyi or follow us on Twitter.

WeVisOne-VL connects box predictions to pixel masks

WeVisOne-VL builds on Qwen3-VL and uses a special token to connect each predicted bounding box to a pixel-level output. The 2B and 4B models handle detection, referring localization, counting, OCR, segmentation, and matting through structured next-token prediction, with outputs ranging from points and boxes to masks and alpha mattes.

The central design is a lightweight pixel decoder with about 9.6 million parameters. It combines the hidden state of each generated token and its predicted box with multi-level features from the shared vision encoder, then produces a full-image mask or alpha matte. This lets one VLM produce discrete locations and dense instance-level predictions without a separate large segmentation model. The decoder's use of intermediate vision features also relates to questions about visual information processing inside VLMs, though this paper presents an architecture rather than an interpretability study.

The authors report evaluation across more than 30 benchmarks and describe leading performance on object- and pixel-level tasks, alongside rejection of queries whose targets are absent. The available paper text gives no numerical headline score for those claims, so the results should be read alongside the benchmark-specific tables rather than as a single aggregate win....

Continue reading the full paper summary on AIModels.fyi →