Restore Face (CodeFormer)

How it fits

Typical backends

Caveats

• I/O contractThe output is a downstream-resolved oneof t[Image, UInt64, [Image]] of the same restoration: the full restored frame, the count of restored faces, or the list of per-face 512x512 crops. The count always equals the crops length. Output pixels are RGB regardless of the input colour order.
• FallbackWhen no face is detected the Image arm returns the source enlarged by upscale, the UInt64 arm `0`, and the [Image] arm an empty list; a zero-sized frame is passed through on the full-frame arm.
• Accuracyfidelity_weight trades restoration aggressiveness against identity: `0.0` enhances most freely and may drift from the original face, `1.0` stays closest to the input; `0.5`-`0.7` is typical.
• Parameter interactionupscale enlarges the whole output frame by that factor on top of face restoration, so a larger factor raises memory and time roughly with the output area; the per-face crops stay 512x512 regardless.
• Costface_upsample and background_upsample each add a Real-ESRGAN super-resolution pass (faces, background) and raise compute and memory; leave them off when only face cleanup is needed.
• CompatibilityThe CodeFormer generator checkpoint is bound through the model file slot; the retinaface detector and face-parsing nets are fetched from the public release host on first use, so a cold container on a restricted network must reach them or pre-seed the cache. A GPU is strongly preferred; CPU restoration is slow.