Segment Referring Expression (Florence 2)

How it fits

Typical backends

Caveats

• I/O contractThe model emits polygon contours; the [Mask] and [Segmentation] arms rasterise each referent's contours into one boolean mask, while the [Polygon<Double>] arm returns every contour separately. The Segmentation arm tags the referent with class id zero and full confidence, since this task has no class space or score.
• AccuracyThe task localises a single described referent per phrase; an expression with no visible match returns an empty list, and an empty phrase likewise yields an empty result rather than an error.
• FallbackAny inference failure for one message is logged and an empty result is returned, so the container stays alive across transient errors.
• CostThis is a sequence-to-sequence vision model with multi-beam decoding, so it is slow for live video; a GPU is strongly preferred and CPU is impractical for streaming.
• State lifetimeThe model and the compute device are captured once at startup; weights download from the Hub on first start and changing the model requires a redeploy.