MiniMax H3 Pass Frames

---

When a generated H3 shot continues from the shot before it, it has to start somewhere concrete - and "somewhere" is this node. MiniMax H3 Pass Frames reads the final frames of a decoded pass back off disk and hands them to the next generation as its keyframe. It's one of the internal seam nodes the Creator's "chained" mode builds into the graph; you won't place it by hand.

The clever bit is where those frames come from. The pass was written to disk the moment it decoded - that's the whole memory strategy of the pack's Reel node - so this node doesn't need a tensor somebody kept alive. It reads the seam's width out of the spill: one frame, or a feathered run of at most 39 for a blended continuation, however long the pass is. That's the mechanism behind the "5/22/39-frame blend" option you see on the Creator's seam picker: the next shot doesn't just inherit one frame, it inherits a run of them, carrying real motion and phase-locked sound across the cut.

The frames come back as 8-bit, which is what the spill stores and what the file was always going to be written as anyway. Worth knowing if you're tempted to worry about fidelity: that's the same diet a keyframe attached from a PNG has always had, and it's the VAE encoder's ordinary food. Nothing is lost that wasn't already gone.

The inputs that matter

• source - the pass to read from, as the Reel node wrote it (the pack's MMC_PASS type). This is the only required input.
• count - optional, default 1. How many frames the seam inherits. A feathered seam gets a run instead of a single frame; the practical ceiling in the blend picker is 39.

Output is a single IMAGE - the frames the next segment continues from. You'd wire it into the next segment's prev_image input if you were building this graph by hand, which you're not, because the Creator does it for you.

The sibling nodes

There's a whole family of these seam helpers. MiniMaxH3PassAudio does the same job for the soundtrack's tail (a single sample says nothing about a room, so sound gets a stretch of seconds instead of one frame), and the MiniMaxH3ClipFrames / MiniMaxH3ClipAudio pair do the equivalent for a supplied clip - footage you attached - rather than a generated pass. The distinction matters because a clip was never generated, so it has no spill to read back; the clip-side nodes decode what they need on demand, and only when something actually continues from them.

Installing

Ships with the pack, dev-only, no separate install:

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

Restart ComfyUI, no pip install, standard H3 weights. If you're hand-building graphs and want to see what a chained render is actually doing, this is one of the few internal nodes worth peeking at - it's where "the next shot continues from the last one" stops being a promise and becomes a list of pixels.