Analyze Face (FaceXFormer)

How it fits

Typical backends

Caveats

• I/O contractThe output arm is fixed by the downstream connection: every head is computed in the same pass, so the resolved type only chooses which result is surfaced — the Image arm returns a single colorized parse map for the whole frame, while the list arms fan out one or more entries per face.
• Hard constraintA face box that expands or crops to a degenerate region is silently dropped, so the number of emitted entries can be fewer than the number of input boxes; the metric arm instead keeps positional alignment by emitting an empty bag for a dropped face.
• Parameter interactionbbox_margin widens each input box before cropping, which gives the backbone more facial context; too small a margin clips the hairline and jaw and degrades parsing and landmark accuracy.
• Fallbackdevice requesting a GPU on a host without one falls back to CPU, where per-face latency rises sharply because the backbone runs once per face crop.
• AccuracyDemographic, attribute and head-pose labels follow the public training datasets the checkpoint was trained on and inherit their distribution biases; treat them as soft estimates rather than ground truth.
• LatencyWeights download and load once at startup, so the first call on a fresh container pays the model-load cost on top of inference.