avatar

Tencent open-sources WeVisOne-VL with absent-object rejection

Tencent's Apache 2.0-licensed WeVisOne-VL models handle detection, counting, OCR, segmentation and alpha matting, and return None when an object isn't in the image.

avatar
Surf AI ·

TL;DR:

  • WeVisOne-VL comes in 2B- and 4B-parameter versions under the Apache 2.0 license.
  • On Tencent's WeRef benchmark, the 4B model rejects 96.4% of absent-object queries; on Ref-ZOM, 86.0%.
  • The image-only models combine five vision tasks in one set of weights and use a 9.6-million-parameter decoder without a Segment Anything Model dependency.
  • Explicit absent-object rejection can reduce fabricated detections in inspection, visual search and automated annotation workflows.

Headline

Tencent has open-sourced WeVisOne-VL, which combines five image-understanding tasks and explicitly rejects absent-object queries.

Summary

Tencent released Apache 2.0-licensed 2B- and 4B-parameter WeVisOne-VL models. They can return None when asked to locate an object absent from an image, rather than return a bounding box. Tencent reports that the 4B model rejects 96.4% of absent-object queries on its WeRef benchmark and 86.0% on the public Ref-ZOM benchmark. The image-only release also combines detection, counting, OCR, segmentation masks and alpha matting in one set of weights, with a 9.6-million-parameter decoder and no Segment Anything Model dependency.

Analysis

The main advance is treating an absent object as a valid grounding result. That addresses a failure mode where visual-model outputs can turn into false downstream actions. It matters for inspection, visual search and automated annotation, where a fabricated detection can cost more than a missed object. Rejection performance therefore needs to be judged alongside detection accuracy on objects that are present. Combining five tasks in one model could simplify deployment and maintenance for developers who would otherwise integrate separate vision components. Alpha mattes also provide soft opacity boundaries for hair, fur and transparent edges. Apache 2.0 licensing and published evaluation scripts lower barriers to commercial adoption and benchmarking. The release is a practical alternative to fragmented vision pipelines rather than another general-purpose vision-language model.

Impact Assessment

Surf provides information only and does not constitute investment advice. All investment decisions are made at your own risk. Please conduct your own research (DYOR).