Victor Kuznetsov 3d00fed00c fix(photomaker-v2): compute id_embeds via FaceAnalysis2 before pipeline call ...

The Modal cert sweep against V2 hit the next layer of the API:

  PhotoMakerIDEncoder_CLIPInsightfaceExtendtoken.forward() missing 1 required
  positional argument: 'id_embeds'

V2 forward takes BOTH the CLIP image embedding (computed inside the pipeline from
input_id_images) AND an ArcFace identity embedding (id_embeds) that the caller
must compute. The upstream pipeline does NOT auto-compute it -- inference_pmv2.py
shows the caller using FaceAnalysis2 + analyze_faces to extract the ArcFace
vector from each input ID image and passing id_embeds=torch.stack([...]) into
pipe(...).

Wired the same flow here:
- New _get_face_analyser() singleton (double-checked lock) builds
  FaceAnalysis2(['CUDAExecutionProvider' | 'CPUExecutionProvider']).prepare(...).
  This is the non-commercial step (antelopev2/buffalo_l auto-download on first
  use). Module docstring already calls it out.
- Per face: analyze_faces() -> torch.from_numpy(embedding) -> .unsqueeze(0) to
  match the pipeline's expected (B, D) shape, casting to pipeline.device/dtype.
  Faces InsightFace can't detect inside the crop get skipped (the most likely
  cause would be the diffusion-cleaned face being too small or stylised after
  the main pass; YuNet already gated us into having a face per crop, so this
  should be rare).
- id_embeds= keyword threaded into the pipeline call site alongside the existing
  input_id_images=.

Tests untouched (the V1-only safety guard was already removed in the previous
commit when we swapped V1->V2; the existing 11 tests still pass).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

2026-06-08 18:49:10 -07:00

..

remove_ai_watermarks

fix(photomaker-v2): compute id_embeds via FaceAnalysis2 before pipeline call

2026-06-08 18:49:10 -07:00