MiniMax H3 Pass Audio

MiniMax H3 generates picture and sound as one thing - dialogue, room tone and SFX come out of the same omni-modal latent. That's great, but it makes seamless multi-shot renders subtler than "last frame, next frame," because a single audio sample says nothing about a room. MiniMax H3 Pass Audio is the sound-side seam: it takes the last few seconds of a decoded pass's soundtrack so the next pass can continue from it.

It's an MiniMax/internal node - the Creator/Timeline writes it into the graph during a render; you won't find it by browsing the menu. It's the counterpart to Pass Frames (the picture's seam is one frame; sound's is a stretch, because ambience lives in time, not a single instant).

How it works

The pass lives on disk - the reel node spilled it to temp the moment it decoded - so this node reads a slice of its soundtrack back without ever holding the full pass in memory. It pulls the tail of the audio and hands it to the next segment's prev_audio input, where it becomes the starting state the next shot's audio continues from. The default length is deliberately short (1 second): just enough to carry the room's character across the seam without shackling the next shot to the previous one's sound.

Inputs and output

• source - the MMC_PASS handle to the decoded pass. Required.
• seconds - optional float, default 1.0, range 0.1 to 4.0. How much of the tail to take. Turn it up when a scene needs real continuity (a voice carrying across a cut); leave it low for a hard scene change.

Output is a single AUDIO tensor, ready for the next segment's prev_audio slot or any audio input in the graph.

The honest trade-off

The 1-second default is a judgment call, not a limitation: a longer tail gives the next pass more to latch onto but also more of the old sound to fight against, and H3's jointly-generated audio means the tail and the new pass have to agree or you hear the seam. If your audio seams feel like the ambience "jumps" between shots, the first knob to touch is this one - but a second knob, on the Timeline's transition settings, controls how the tails crossfade in the final mux. Same pack, different node.

Install is the shared one: ComfyUI Manager → search "H3 Creator Palette", or git clone https://github.com/z3rofeels/ComfyUI-H3-Creator-Palette into custom_nodes/, restart, hard-refresh. ComfyUI 0.34.0+ required. If a hand-built graph hands this node an empty source, expect a clear error - the pack would rather tell you the pass is missing than feed the next segment silence.