ComfyUI-ALLinONE-MinimaxH3

found

Description compacte indisponible pour ce nœud.

H3AudioJoinSmooth

Fades the audio seam at the extend join. Extend output is [source audio] + [generated audio], but the generated audio continues the source tail through a lossy VAE round trip, so the two sides rarely meet sample-for-sample and the hard concat clicks. This node overlaps the source tail and the generated head with a short linear crossfade (the two sides are the same audio continuing, so an equal-power blend would boost the middle), then re-pads to the exact frame-derived duration so audio and video stay in sync. The crossfade is deliberately short (0.25s default) so speech and music keep their rhythm.

One Node · entrées 3 · paramètres 3 · sortie 1

H3AudioTrim

Trims an AUDIO input to a target duration (clamped to the 15s H3 ref spec). MiniMax H3 reference audio clips are specified as 2-15 seconds, and an audio ref longer than the target video is out-of-distribution for the packed layout - this node keeps ref audio within spec.

One Node · entrée 1 · paramètre 1 · sortie 1

H3CacheBust

Internal cache-invalidation node (inserted by the JS between the CLIP loader and the H3 conditioning node). Why it exists: ComfyUI's execution cache fingerprints each node from its input values AND the fingerprints of the nodes wired into it. V3 autogrow inputs (ref_images / ref_audios / ...) arrive as dicts of links - and the cache does not traverse links nested inside dicts. A changed reference image/audio therefore left the cache signature unchanged and every downstream node (conditioning -> sampler -> save) was served stale output. This node sits upstream of the conditioning node and returns a digest of every input that must invalidate generation: the prompt, all media file names, plus the on-disk content of those files (so replacing a file under the same name also invalidates).

One Node · entrée 1 · paramètre 1 · sortie 1

H3IdentityAnchor

Pins a reference image as a stock first/last keyframe anchor on the H3 conditioning, so the shot starts (or ends) exactly on that image. Uses only core ComfyUI keyframe support - no third-party layout patches required.

One Node · entrées 4 · paramètres 4 · sortie 1

H3MaskVideoPrepare

Prepares mask-mode video inputs at 24 fps by selecting frames, preparing source audio to match the frame-derived duration, and producing stereo 32 kHz model audio. Returns the prepared images, model audio, source audio, fps, and frame count; rejects target frame rates other than 24 fps.

One Node · entrées 2 · paramètres 3 · sorties 5

H3MotionRefScale

Build the H3 motion reference from the tracked source crop. The ref video only supplies movement, never identity, so it can be small without hurting the replacement. The degrade mode picks how the tracked subject is hidden from the footage that rides into the H3 conditioning so the original person cannot copy itself into the regenerated region: - source: the real crop passes through (most faithful motion, but the original identity can bleed into the start of an edit). - silhouette: the subject is painted as a plain white shape over the real scene, leaving only coarse pose. - chroma: the subject's color is replaced with blocky noise while its luminance is kept, so articulation and shading survive but the original face, skin and clothing colors are gone. Keeping the video on the 32 grid means the H3 node's own canvas logic leaves it at the reduced size.

One Node · entrées 2 · paramètres 2 · sortie 1

H3OneNode

Output node that returns a poster image and relative video path from the most recently recorded output for the node. It attempts to create a poster when the output file exists and returns an empty image when no poster is available.

One Node · entrée 1 · paramètres 0 · sorties 2

H3OneSAM3CropCheck

Hands the frontend a JSON report of the crop MVEx Subject Crop plans around the tracked subject plus SAM3's own per-object confidence, so the tracking preview can draw the crop box and flag shaky tracks before an H3 run. Purely additive: it only reads the mask pipeline and writes a UI string, so wiring it into a real generation graph cannot change the output.

One Node · entrées 3 · paramètre 1 · sortie 1

H3PaintedRegion

Follows a user-painted whole-head region through the video using the SAM3 track, so hair and accessories stay attached to the head instead of sitting at their first-frame spot. Each frame the painted mask is shifted by the track's centroid drift since frame 0, then OR-ed with that frame's track mask. A track gap falls back to the last shift so the region never jumps or vanishes. The painted mask is one frame; the track is one frame per video frame.

One Node · entrées 2 · paramètre 1 · sortie 1

H3StitchFrames

Internal side-by-side compare stitcher for the Compare & Stitch page. N IMAGE inputs (1-4) become a single IMAGE batch arranged in a grid, so a plain VHS_VideoCombine can write the comparison clip without any external stitch pack. Frame matching is done upstream by the VHS loaders; this node only pads shorter clips and lays the grid out.

One Node · entrées 4 · paramètres 4 · sortie 1

ok

Description compacte indisponible pour ce nœud.

SomeOtherNode

Description compacte indisponible pour ce nœud.