ComfyUI Node
MiniMax H3 Audio Preview
Hear the shot while it denoises. A model like MiniMax H3 generates stereo audio in the same forward pass as the picture, and ComfyUI's preview discards it: prepare_callback keeps x0.tensors[0] and drops the audio stream, and the preview socket has no audio event to carry it anyway. This node decodes that stream itself and draws a waveform you can play. It attaches alongside whatever draws your frames rather than replacing it — KJNodes' Preview Override with taeh3, or any other. Wire it anywhere between the model source and the sampler.
MiniMax H3 Audio Preview
- model
- audio_vae
- MODEL
◄window_seconds3.0►
◄start_at_percent50►
◄every_n_steps1►
◄max_preview_overhead15►
CategoryMiniMax H3
Inputs (6)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | Model to attach the audio preview to. | |
| audio_vae | VAE | minimax_h3_audio_vae. The real decoder, and the one that says how many channels the audio stream has so it can be picked out of the pack. H3's audio has no tiny equivalent of taeh3, and does not need one: a three-second window is ~110 MB and a fraction of a second. | |
| window_seconds | FLOAT | 3.00–15 | Seconds of audio to decode, from the START of the shot — the part that lines up with a preview animation looping from frame one. Cost scales with it. 0 decodes the whole clip (~520 MB at 15s). |
| start_at_percent | INT | 500–100 | Don't decode until this share of the steps is done. Early on the audio stream is still mostly noise, so the decodes would be spent on hiss. |
| every_n_stepsopt | INT | 11–50 | Never update more often than every N sampler steps. |
| max_preview_overheadopt | INT | 150–100 | Cap on how much of the render time this may use, in percent. Updates are also never sent faster than the clip plays, so you always hear one through. 0 disables the cap. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | Model with the audio preview attached. |