Locate Objects (Eagle LocateAnything)

How it fits

Typical backends

Caveats

• I/O contractThe scene Image and the description String arrive together; the description conditions what is located. An empty description falls back to a generic locate-every-object prompt so the component always answers.
• I/O contractThe output is a downstream-resolved oneof, fixed once the graph is deployed: BoundingBox boxes carrying a class id and confidence, bare Rectangle geometry, boxes paired with the grounded text label (`{bounds, label}` -- the only arm where the open-vocabulary text survives), a UInt64 instance count, or an Image overlay with the boxes drawn. prompt_to_class_id maps grounded labels to class ids for the BoundingBox arm; unmapped labels default to class id `0`.
• FallbackAn empty image, a no-match answer, and any caught inference error all degrade to the resolved arm's empty value -- an empty list, a zero count, or a blank overlay frame -- rather than crashing the worker; every input still produces exactly one output.
• AccuracyCoordinates are decoded as integers over a 0..1000 grid and rescaled by the input extent, so boxes are quantised; the model grounds real-world imagery best and answer quality drops on abstract or heavily distorted scenes.
• LatencyThis is a multi-billion-parameter vision-language model; CPU inference is impractical and a GPU with sufficient memory is required for live use.
• State lifetimeThe model and processor are loaded once at startup and reused for every call; the device, token budget, and decode mode are fixed at that point and a restart is needed to change them.