Ground Phrase (Florence 2)

How it fits

Typical backends

Caveats

• I/O contractThe phrase is a per-frame String; an empty phrase or an empty Image yields an empty result rather than an error, so silent or unframed ticks produce zero regions.
• AccuracyGrounding only finds noun phrases that the supplied caption mentions, so a phrase describing something absent from the frame returns nothing; pair it with a captioner to ground generated text or feed an operator phrase directly.
• I/O contractLabels are free-form phrase fragments with no fixed class-id space, which is why the labelled arm emits `{bounds, label}` records rather than enumerated boxes; the geometry-only arm drops the labels.
• FallbackInference failures on a single frame are caught and logged, returning an empty result for that frame instead of crashing, so transient errors degrade to gaps rather than a halt.
• LatencyThis is a beam-searched sequence-to-sequence vision model; CPU execution is impractical for live video, and the device binds once at startup so it cannot be switched at runtime.