Segment Image (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 instances, so masks, boxes and classes stay aligned. The instance list is unbounded, so a common concept on a crowded scene can return many results per frame.
• Hard constraintprompt must be a non-empty phrase, or a comma-separated list of non-empty phrases, at startup; an empty value fails to start. An invalid runtime update is rejected and the previous prompt is kept.
• LatencyMultiple comma-separated phrases in prompt are segmented one after another, so latency grows with the number of phrases per frame.
• Parameter interactionprompt_to_class_id assigns a class id per phrase so downstream arms can distinguish them; any phrase absent from the map falls back to class id zero.
• CompatibilityThe SAM 3 checkpoint is provided through the model weights slot, so the component runs without runtime network access. Weights load in half precision on GPU to fit a single device; device falls back to CPU (full precision) with a warning when no CUDA runtime is present.
• State lifetimeEvery setting except prompt is captured once at startup; prompt is re-read on each frame so it can be steered at runtime, while model and device need a redeploy to change.