MiniMax H3 Clip Audio

A generation running into a clip shouldn't slam to a stop on the footage's first frame with dead air after it. MiniMax H3 Clip Audio grabs the first or last seconds of a supplied clip's soundtrack so the generated shots around it can inherit its sound - the audio counterpart to Clip Frames, and the footage-side twin of Pass Audio.

It's an MiniMax/internal node, placed by the Timeline when a shot sits next to a footage card, not something you'll find by browsing the node menu. But it's the node that makes a render with supplied footage feel like one continuous piece instead of a patchwork.

How it works

Like its siblings, it works from clip_data - the JSON spec naming the file - not from a tensor. The at combo picks which end of the clip's soundtrack you want: tail is what a generation after the clip continues from, head is what a generation before the clip ends on. It reads just those seconds off the file, resampled to the generated passes' decode rate (that's what the audio_vae input is for on the reel side - this node itself just reads the file's own rate). The result feeds a segment's prev_audio or next_audio slot, becoming the starting state the next shot's audio grows from.

Inputs and output

• clip_data - the clip spec. Required.
• seconds - optional float, default 1.0, range 0.1 to 4.0. How much of the clip's sound to take. Same trade-off as Pass Audio: more gives the next pass more to latch onto, less lets a scene change actually change.
• at - head or tail, default tail.

Output is a single AUDIO tensor.

Why the default is short

One second of tail is the pack's default everywhere in this chain, and it's a deliberate judgment call rather than a lazy constant: H3's audio is generated jointly with the picture, so a seam's soundtrack has to agree with both the clip it's leaving and the room the next pass invents. Too long a tail and the new shot is narrating over the old room; too short and you hear the cut. If a seam sounds "wrong" but you can't place why, try nudging seconds before you touch any sampler setting - it's the cheapest experiment on the board.

Install is the shared pack install: ComfyUI Manager → search "H3 Creator Palette", or git clone https://github.com/z3rofeels/ComfyUI-H3-Creator-Palette into custom_nodes/, restart, hard-refresh. Needs ComfyUI 0.34.0+. If a hand-built graph asks for the head of a clip that has no audio track, you'll get an error, not silence - which is the pack's whole policy on these seams.