MiniMax-H3 Multishot — chained shots, one master, with audio — v1.0 — v1.2

Measured on an RTX 5090, 243 frames, one anchor at pixel frame 121: the rendered frame most resembling the anchor image was frame 122 — the requested position, off by one — reached by continuous motion with no cut (peak frame-to-frame change 2.3× the median), and the audio ran unbroken straight through it.

The patch is applied in memory . It does not edit any ComfyUI file. It self-tests against the stock formula before committing and rolls itself back if first/last positions do not reproduce exactly — so a future ComfyUI change degrades to “interior anchors unavailable” rather than to broken renders.

It moves, or it cuts — and your images decide which

Anchor images with a plausible camera path between them (same place, different angle or framing) make H3 interpolate : a real move that arrives on time. Images with no possible path — a kitchen and a diner — make it cut , then hold.

That is not a limitation of this node. Stock first/last does exactly the same thing with such a pair; it is the model doing the only sensible thing. And the cut case has its own use: a timed shot change inside a single generation , which means the audio runs continuously across it, with no per-shot voice re-derivation and no seam.

Reference-to-video groundwork

The two nodes below are what reference workflows need. The hard-mode graph itself is built but has not been rendered yet, so it is not in this release — it ships when it has a render behind it.

References and keyframes are mutually exclusive , in ComfyUI core rather than by choice here. model_base.py assigns cond_video_latents for refs, discarding any keyframe latents while the keyframe layout rows survive — the packed sequence then desyncs into a shape-mismatch crash. There is no “reference images plus start frame” mode. Pick one per shot.

Two things that are easy to get wrong and are documented on the canvas: reference tags are 1-based in the prompt ( ) while the input slots are 0-based ( ref_image_0 ); and a reference video with a soundtrack consumes an ordinal before your standalone clips.

Two more nodes

H3 Reference Audio (stereo guard) — the audio VAE encodes [B, 2, L] and the layout reserves exactly two channels, so a mono reference produces half the rows it reserved and dies deep inside the model with no useful error. Nothing in stock converts it. This forces stereo 32 kHz and trims length.

H3 Condition Strength — exposes minimax_visual_cond_noise_aug and minimax_audio_cond_noise_aug , which core reads but no stock node writes, so they sat at 0.999 / 1.0 permanently. They blend noise into the conditioning latents and set the timestep the condition rows occupy: an anchor-strength dial.

Still in the box from v1.1

Image-to-video via an optional start_image .

~4× faster on 32 GB cards. The text encoder is evicted before sampling — encoder (~16.5 GB) and DiT (~25 GB) do not co-fit, so the DiT was loading partially and streaming ~19 GB from system RAM every step. RTX 5090: ~60 min to ~15 min, same render. The keyframes node does this too.

Long-form memory sampler for 2–5 minute pieces.

Audio seams crossfaded; seed_per_shot on by default.

Links

Node pack on GitHub

H3 DiT GGUF

Text encoder GGUF (+ mmproj)

VAEs

Support

Everything I publish is free and stays free. If it saved you a night of debugging, tips keep the 5090 warm: Ko-fi · GitHub Sponsors · Liberapay .

Text encoder · Unknown