| name | dub-casting-verify-content-type-from-frame |
| description | Use when choosing the dub-studio casting content_type (real vs anime/cartoon) or a face-detection path, or when picking casting models for a video — verify the ACTUAL footage, never infer live-action-vs-animation from the show's title. |
Casting content_type: look at a frame, don't guess from the title
dub-studio casting has two paths: content_type="real" (SCRFD det_10g.onnx + LVFace LVFace-L_Glint360K.onnx, cosine) for photographic human faces, and content_type="anime" (anime_face detector + CCIP) for drawn/cartoon characters.
Never pick the path from the title. Many franchises exist as BOTH animation and live-action — e.g. Avatar: The Last Airbender is a cartoon AND a Netflix live-action series with real actors. Assuming "anime" from the name and reaching for the anime models on live-action footage is wrong (the user's live-action Avatar test is real).
How to apply
- Before choosing content_type, sample a frame (
ffmpeg -ss <t> -i in.mp4 -frames:v 1 f.png) and look at it — real photographed faces → real; drawn/animated → anime.
- Model paths resolve directly under
<repo>/models/faces/, NOT in scrfd//lvface/ subdirs: models/faces/det_10g.onnx + models/faces/LVFace-L_Glint360K.onnx (real), models/faces/anime_face/model.onnx + models/faces/ccip/model_feat.onnx (anime), plus models/faces/wespeaker/… (voice) and models/faces/lr-asd/… (active-speaker). Check the real file path before concluding a model is "missing".
- SCRFD/LVFace are NOT in the download manifest (see setup.rs) — they're placed manually; if absent, casting falls back to voice-only (no avatars).
Related: [[feedback_never_guess_verify_only]], [[handoff_casting_115_2026_07_18]].
