ComfyUI Node
H3 Audio Drive Auto (align + lock)
Aligns an audio track to MiniMax H3's 17k+5 frame grid and freezes it into the joint AV latent (video denoises, audio is masked at denoise 0). Returns the duration-aligned audio for an exact final mux, the frame-grid-legal length, and a human-readable report.
H3 Audio Drive Auto (align + lock)
- audio
- audio_vae
- av_latent
- av_latent
- audio_aligned
- length
- report
◄modelock_source►
◄remix_strength0.35►
◄pad_modesilence►
◄width1344►
◄height768►
◄max_window_seconds14.0►
◄window_overlap_seconds0.5►
◄audio_offset_frames0►
Categoryminimax/h3/audio
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| audio | AUDIO | Voice/dialogue that should drive the mouth. The same clip is returned unchanged (duration-aligned) for the final mux. | |
| audio_vae | VAE | MiniMax H3 audio VAE (minimax_h3_audio_vae_fp32.safetensors). | |
| mode | COMBO | lock_source | lock_source: the source audio is frozen into the latent and only drives the video. remix: audio is partially denoised (keeps timing, lets the model re-voice it). native: leave the audio half free so the model invents its own track. |
| remix_strength | FLOAT | 0.350–1 | Only used when mode=remix_source. 0=fully locked, 1=fully regenerated. |
| pad_mode | COMBO | silence | When the audio is shorter than the video: silence-pad, trim the video length down to the audio, or raise an error. |
| width | INT | 134432–16384 | — |
| height | INT | 76832–16384 | — |
| max_window_seconds | FLOAT | 14.01–60 | Long-audio planning only. H3 was trained on ~124-362 frames (5-15 s); longer tracks should be rendered as windows. |
| window_overlap_seconds | FLOAT | 0.50–5 | Overlap suggested for a windowed render (continuity guide). |
| audio_offset_frames | INT | 0-48–48 | Shift the RETURNED audio against the rendered video, in frames at 24 fps (length preserved). H3 tends to play the mouth ~10 frames ahead of the track, so a value of -10 lands the two on the same frame; use the offset H3LipSyncScore reports. The latent that drives the performance is never shifted. |
| av_latentopt | LATENT | AV latent from MiniMax H3 Image/Reference to Video. If connected, the audio is aligned to THAT latent. If omitted, a clean empty AV latent is built from width/height/audio duration. |
Outputs (4)
| Name | Type | Description |
|---|---|---|
| av_latent | LATENT | — |
| audio_aligned | AUDIO | — |
| length | INT | — |
| report | STRING | — |