MiniMax H3 Pass Audio

---

Video seams have a picture problem and a sound problem, and they're not the same problem. The picture side - what the next shot continues from - is a single frame, handled by MiniMaxH3PassFrames. The sound side is this node, and the pack's own docstring nails why it can't work the same way: a single sample says nothing about a room. A room is defined by its reverb tail, its ambience, the way dialogue bounces. So MiniMax H3 Pass Audio hands the next generation a stretch of the previous pass's soundtrack - the last few seconds, by default one - and that's what the next shot's sound continues from.

This is the mechanism behind the "phase-locked sound across the cut" claim in the Creator's chained-mode seam options. When a seam continues, it doesn't just carry pixels; it carries the audio context, so a conversation doesn't change rooms between shots and the room tone doesn't suddenly restart. The audio half of the H3 latent is generated jointly with the picture - that's the whole omni-modal point of H3 - so giving the next pass the previous one's sonic tail is what keeps a cut feeling like one continuous place instead of two clips stuck together.

It's internal plumbing, dev-only, written into the graph by the Creator when you pick a chained seam. You won't place it by hand, and its only custom input is the MMC_PASS reference.

The inputs that matter

• source - the decoded pass whose soundtrack you're reading, as the Reel node wrote it.
• seconds - default 1, range 0.1–4. How much of the tail the next pass inherits. The default is short on purpose - enough to carry a room, not enough to drag the previous shot's whole soundscape along.

Output is an AUDIO tensor, the tail you'd feed to the next segment's prev_audio input - which the Creator wires for you.

The family it belongs to

This node and MiniMaxH3PassFrames handle seams between generated passes. The MiniMaxH3ClipAudio / MiniMaxH3ClipFrames pair do the equivalent for supplied footage - a clip you attached, which was never generated and so has no spill to read back. The clip-side nodes take a clip_data blob and an at switch (head or tail) instead of a pass reference, and they only decode what they need when something actually continues from the clip - a clip that nothing continues from is never decoded at all.

Installing

Ships with the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/roadmaus/ComfyUI-MiniMax-Creator

Restart ComfyUI, no pip install. It needs the usual H3 weights behind it and a ComfyUI that ships comfy_extras/nodes_minimax_h3.py. If you've ever had a multi-shot H3 piece where the cut visually lands but the audio audibly resets the room, this is the node doing the unglamorous work that stops it.