MiniMax-H3 Sections — Unlimited Length, Re-Roll Any Shot v1.0

A sectioned film workflow for MiniMax-H3. The film is built as a chain of sections, each one continuing the previous section's motion AND sound, so a character keeps moving and keeps talking across a cut instead of restarting. Add as many sections as your story needs, and re-render any single one without touching the others.

How it works

Each section is a complete render of its own: prompt, seed, steps, length, LoRAs. What connects them is the latent handoff — a section receives the previous section's raw output latents and locks its opening onto them, picture and audio together, then trims the overlap it regenerated. Nothing is decoded and re-encoded between sections, so quality does not degrade along the chain and the sound crosses the join as continuously as the picture.

Because each section is a separate set of nodes rather than one big render, they are cached independently. Change section 4's prompt and queue again: sections 1-3 are already computed and cost nothing, so only 4 onward re-renders. That is the whole reason the workflow is shaped this way. Tuning a long film normally means paying for the entire thing every time you fix one shot; here you pay for the shot you changed and the ones after it.

Six sections ship set up. They are a starting point, not a limit — sections are copies of each other, so you extend the film by copying the last one, retitling it, and redrawing three links. The on/off panel picks up the new section by itself. That is also how the workflow steps past H3's single-pass ceiling of 362 frames: no individual render is longer than the model was trained for, but the film can be as long as you like.

What it is good at

• Long films with dialogue. Continuous speech across cuts is what this chain does best, and it is what a single long render cannot give you.

• Iterating cheaply. Re-roll one beat and the rest come back from cache. An assembled film can also be rebuilt in seconds at any point, since assembly does no sampling — useful for watching the story so far while you write.

• Uneven films. Length, steps, seed and LoRAs are all per-section, so a hero shot can afford 20 steps and a style LoRA while a filler gets 10 and none.

• Writing rather than wiring. All the prompts sit in one column on the left, one box per beat. You write the film top to bottom and never touch the graph.

• Nothing custom of mine to install. It runs on two H3 packs plus rgthree, all free and linked below - see the install section if ComfyUI says the nodes are missing.

What it does not do yet

Stated plainly, because these are the things you will notice.

• No audio spine. H3 generates the sound and it carries across joins, but you cannot yet supply your own soundtrack or a voice reference for the chain to lock onto. This is the next thing planned, and it is the biggest gap.

• Voice drift is unmeasured. The voice is anchored only by your prompt and reference picture, so across many sections it may wander. If it does, the audio spine above is what will fix it.

• One character. Multiple reference pictures are read as the same person on this path, so two-character scenes are out of reach for now.

• No global colour matching. Every section renders on the same checkpoint and the joins are latent-continuous, so drift is mild — but there is no pass that levels the whole film's tone at the end.

• Forward-only. Changing a section makes the sections after it stale, because they were built on its latents. Fix a beat, then re-roll the ones that follow. Never regenerate backwards.

Requirements

• ComfyUI

• ComfyUI-H3-Multishot (github.com/jlucasmcrell/ComfyUI-H3-Multishot) - the H3 loaders, LoRA stack and A/V concat

• ComfyUI-H3-Motion-Context (github.com/NikoDemon80/ComfyUI-H3-Motion-Context) - Motion Context and Trim, which are what chain the sections together. This one is essential and easy to miss.

• rgthree-comfy (github.com/rgthree/rgthree-comfy) - the Fast Groups Bypasser used as the sections on/off panel

• Models: an H3 ref2va checkpoint (fl2va has no reference rows and will not work here), the H3 text encoder, and both the video and audio VAEs

Installing the two H3 packs (no command line needed)

Neither H3 pack is listed in ComfyUI Manager's search, so looking up node names like H3ConcatAV or MiniMaxH3MotionContext there finds nothing. That is why you may see "custom nodes aren't installed". Two ways to fix it, easiest first.

Option 1 - ComfyUI Manager, no files to move. Open Manager, choose Custom Nodes Manager, click "Install via Git URL", and paste each address below, one at a time. Manager downloads and places them for you.

Option 2 - download the ZIPs. Open each address below in your browser, click the green Code button, choose Download ZIP, then extract it into your ComfyUI\custom_nodes folder. Important: the folder you drop in must contain the pack's files directly. If you end up with custom_nodes\ComfyUI-H3-Multishot-main\ComfyUI-H3-Multishot-main\ then move the inner folder up one level, or ComfyUI will not see it. The -main on the end of the name is harmless.

• https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

• https://github.com/NikoDemon80/ComfyUI-H3-Motion-Context

• https://github.com/rgthree/rgthree-comfy (this one IS in Manager's search, so you can install it the normal way)

Then close ComfyUI completely and start it again - a browser refresh is not enough, the packs load at startup. Reload the page afterwards and open the workflow. One note: MiniMaxH3ReferenceToVideo is built into ComfyUI itself, so it needs no pack; if that one is missing, your ComfyUI is too old.

Quick start

1. In the GLOBAL group, pick your character picture in CHARACTER REF 1 and unmute that node (Ctrl+M) — a muted node passes nothing, so choosing a file is not enough.

2. Check WIDTH and HEIGHT. 736x1280 ships; both must be multiples of 32.

3. Replace the example beats in the PROMPTS column with your own.

4. Queue. Each section writes its own clip; the FILM group writes the assembled film.

Two rules the on-canvas notes repeat, because both are easy to break by accident: keep every seed's control on 'fixed' — one seed set to 'randomize' re-renders its section and everything after it on every queue — and always regenerate forward, never backward.

Three notes on the canvas cover the rest: first-run steps, what every section holds, and how to add another section. Frame counts follow H3's grid of 17k+5 (90, 107, 124, 141 ... 362) at 24 fps, and sections after the first deliver 22 frames fewer than their length, because the regenerated join is trimmed.