Detect Objects (Florence 2)

How it fits

Typical backends

Caveats

• I/O contractFlorence labels are free-form strings with no built-in integer class-id space; the ordered classes config is the catalogue. A detection's label is mapped back to its catalogue index, and that index is the emitted class id; a label matching no catalogue entry is dropped. With an empty catalogue every detection is kept and labels are assigned ids in first-seen order. Florence carries no per-detection score, so each detection is reported at full confidence.
• AccuracyThis is closed-set generic object detection from Florence-2's detection task prompt; to detect a specific named thing use an open-vocabulary or phrase-grounding detector instead.
• LatencyDetections are decoded autoregressively, so per-frame latency is higher and more variable than a single-shot detector, which is impractical for live video — especially if the host falls back to CPU.
• CompatibilityThe checkpoint loads offline from the bound model artifact (a snapshot directory holding config.json, the safetensors weights, and the processor/tokenizer files) with `local_files_only`; `trust_remote_code` runs the snapshot's own modeling code, so no network fetch occurs. Bind a Florence-2 checkpoint to the model slot at deploy.
• State lifetimeThe model, device, and classes catalogue are resolved once at startup; changing any of them requires a redeploy.