New node:
- H3 Load Video + Face Select loads the video, detects every face and every
hard cut in one pass, and settles the subject before the graph runs. Adds a
Pick faces dialog for choosing the subject per shot, and an on-node X/Y
coordinate preview. Emits images, audio, face_pick, preview, report,
frame_count and fps.
H3 Face Track + Crop:
- face_pick input takes boxes, cuts and the chosen subject from the node above
- frame_count output, for the H3 node's length
- absent_shots skips shots the subject is not in
- identity_model selects the identity backend: insightface, clip_vision or ccip
- cut_detection splits the clip into shots via PySceneDetect
- select expanded from two modes to nine, with direction in the name, plus
closest_to_xy taking X, Y and frame_index
- canvas_width / canvas_height default to 768; the auto canvas modes clamp up
to a minimum of 512x512
H3 Per-Frame Denoise:
- takes and returns MODEL; the mask now reaches the result through the sampler
alone and held frames are re-noised to the sampler's current sigma
- strength widgets renamed denoise_multiplier_small_face / _large_face
Example workflows:
- split by how the face is chosen: H3_Face_Refine_Auto_Select.json and
H3_Face_Refine_Manual_Select.json
- both carry the SAM mask pair and the GGUF loader pair muted, with notes
Dependencies:
- scenedetect>=0.7 added, backing cut_detection
Saved 1.0.0 workflows need the model routed through H3 Per-Frame Denoise.
Carasibana
b58be5e807