title: INSID3 with DINOv3: Candidate-Mask Contract description: "INSID3 with DINOv3 transfers a supplied reference mask through a named frozen-backbone release as a candidate segmentation, not ground truth; bind the repository and model revisions, license and access, reference/mask provenance, preprocessing and resolution, positional-bias configuration, uncertainty policy, and source-disjoint review before use." category: reference tags: [segmentation, insid3, dinov3, candidate-masks, provenance, review] aliases: ["In-Context Segmentation with INSID3 and DINOv3", "INSID3 DINOv3 Segmentation"]
INSID3 with DINOv3: Candidate-Mask Contract¶
The upstream INSID3 repository describes a training-free in-context segmentation method operating with one frozen DINOv3 backbone and a supplied reference image/mask. The DINOv3 reference repository publishes the backbone artifacts. Together they justify testing the named release; they do not turn a predicted mask into ground truth, consent evidence, identity evidence, a medical conclusion, or an automatic training label.
Bind the exact release¶
For every candidate mask, retain:
- INSID3 repository and code revision, DINOv3 model/weight identifier, artifact digest, license/access terms, and environment/dependency record;
- reference image and mask digests, author/rights record, class/part definition, crop/orientation/preprocessing, and reference-quality review;
- target asset digest, target-domain identifier, preprocessing, input resolution, color/orientation policy, and output mapping to source pixels;
- documented positional-bias/debiasing, refinement, threshold, and post-processing configuration, including changes from the upstream default;
- output mask, uncertainty/failure signals when available, reviewer decision, correction artifact, and provenance link; and
- source-disjoint validation split and task-specific quality/preservation evidence.
Do not substitute a local adapter, inferred internal feature layout, or monkey-patch for the upstream interface unless that integration has its own versioned implementation and evaluation receipt.
Review candidate masks before use¶
Inspect thin structures, small or repeated objects, occlusion, reflections, transparent materials, blur, shadows, difficult boundaries, and protected regions in the source coordinate system. A plausible boundary can still be semantically wrong or unsafe for the intended edit or dataset.
Keep source assets and derivatives together when splitting evaluation data. The reference image, near-duplicate crops, the same subject, or related derivatives must not leak between tuning and holdout review. Report transfer quality, failure detection, domain shift, reviewer correction, and operational escalation separately.
Failure boundary¶
If the reference/mask provenance is incomplete, the model release or license is unknown, preprocessing differs from the declared run, the target is out-of-distribution, or review cannot validate a protected region, keep the mask in visible review state and request manual annotation. Do not silently accept, relabel, or use it as a source of factual, identity, medical, or release-ready truth.