UNVEIL extends SAM 3 to map concepts like shadows, mirrors and camouflaged objects
A research team reports a trained add-on that lets SAM 3 answer class-agnostic questions, such as where light is blocked or where structure lies, without altering its original outputs.
In plain words
SAM 3 is a vision model that finds and outlines objects it is asked about. The authors say it can recognise many visual concepts internally but often fails to report them, so they trained a companion model, UNVEIL, that turns those internal signals into maps of concepts such as shadows or transparent objects.
What was announced
The paper introduces promptable concept maps, extending SAM 3's promptable concept segmentation from noun instances to maps of concepts named by short noun phrases, with "none" as a possible answer. To evaluate this, the authors built PCM-Bench, which groups the standard test sets of seven dense prediction tasks by the question each asks, covering phenomena such as salient and camouflaged objects, shadows, transparency and mirrors.
UNVEIL trains a copy of SAM 3's semantic path, using LoRA on copies of its last blocks, and adds a separate presence decoder that learns from verified negatives whether a concept occurs. The authors report that a low image-level presence score in SAM 3 can block every detection from its instance path, which UNVEIL addresses. On each of the seven tasks, they say UNVEIL reaches the highest score among compared promptable models, including SAM 3's official fine-tuning and pipelines built on multimodal LLMs. It runs next to SAM 3, leaving SAM 3's own outputs unchanged, and the same weights also map thin structures such as edges and skeletons. Code is available on GitHub.
What it does not show
- The findings come from the paper's abstract only; the full article's methods and details were not reviewed.
- The reported scores are the authors' own comparisons on PCM-Bench and have not been independently confirmed.
For security and IT teams
- Scope
- The released code covers UNVEIL, which runs alongside SAM 3 without changing its outputs, per the paper's abstract.
Sources
Source-support review completed 21 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 21 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.