Segment Image Everything (SAM 3)

How it fits

Typical backends

Caveats

• I/O contractThe output arm is a single downstream-resolved choice fixed at deploy time; every arm reprojects the same auto-generated instances, so masks and boxes stay aligned.
• I/O contractAutomatic masking is class-agnostic: there is no semantic label, so the Segmentation and BoundingBox carriers report the unlabeled instance id zero and use the SAM 3 mask-quality score as the confidence. There is no class-bearing arm.
• AccuracyThe predicted-quality threshold gates mask quality; raising it returns fewer, higher-confidence masks and lowering it surfaces more marginal ones.
• CostThe points-per-batch setting trades latency for memory: more points per batch run faster but use more GPU memory, and the result list is unbounded so a busy scene yields many instances per frame.
• CompatibilityThe checkpoint is bound through the model bundle file slot, not a config param. The gated weights are fetched under the operator's Hub credentials at the fixture stage / first run; once cached the component can run offline.
• State lifetimeThe model binding and all configuration are captured once at startup; changing any setting requires a redeploy.