H3 Video Outpaint · Serial Sample (EXP)

Long-video generation in ComfyUI has a well-worn shape: slice your clip into chunks, generate chunk one, feed it forward, generate chunk two, and spend your evening clicking queue. MiniMaxH3VideoOutpaintSampleT8 is the "do all of that in one queue" node - it walks every planned window serially, inside a single execution, and commits each one to disk as it finishes.

Two routes in this pack lead to a finished outpaint. The candidate route (Candidate → Select → Continue → Compose Candidate) puts a human review gate before the expensive part. This node is the other one: the straight Plan → Prepare → Sample → Compose chain used by Four_Stages_EXP and its folded-up Compact_EXP sibling. Same algorithm, same defaults, no gate. If you already trust your geometry and prompt, this is fewer clicks. If you don't, use the candidate route and thank yourself later.

What it takes

Required: model, prepared, seed, steps, resume. The prepared handle must come from the ordinary Prepare node - a regional conditioning provider from the guided route won't work here, and that mismatch fails rather than degrading.

seed defaults to 20260808 and has control-after-generate, so incrementing it in the UI does what you'd expect. steps defaults to 20, which is the native Stock20 path - not the eight-step distilled route. Don't drop it to 8 hoping for OpenVDN speeds; this sampler has no turbo adapter and the compatibility matrix marks that combination unsupported.

resume is the one that earns its place. Any window already committed in this run's sampling cache gets reused instead of re-sampled, so an interrupted or crashed job continues from where it stopped. The pack's own validation includes a real cancellation/restart test on hard-cut material where a fresh process resumed a two-window cache, and both window hashes matched an uninterrupted control run. That's a stronger claim than most pack READMEs make, and it held.

The optional noise_algorithm combo offers t8.outpaint.coordinate_noise/v1 (default) and t8.outpaint.native_cpu_noise/v1. The shipped workflows run Stock20 native noise; there's no reason to switch unless you're deliberately reproducing something.

Outputs are sampled - the handle Compose takes - and sampling_report, a JSON string describing what actually got sampled.

How it's structured underneath

The plan already split your clip into shots (from cut_frames_json) and each shot into windows, so the sampler never carries context across a hard cut. Windows are generated front to back and committed to the run's cache as complete assets rather than as one giant in-memory batch. resume and the correct output both fall out of that design; so does a failure mode, which is that a half-written window cache under a name you've since changed settings on won't match the identity check.

Expect it to be slow, and expect the audio to survive. The original soundtrack is carried through the prepare cache and only gets re-attached at delivery - the sampler doesn't invent a new one.

Install

Manager: search MiniMax H3 Audio T8, then restart ComfyUI completely. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

The pack's requirements.txt intentionally installs nothing, to keep your ComfyUI Torch/CUDA stack intact - the base nodes use what ComfyUI already ships. This one is not dependency-free, though: the shipped generation workflows need ComfyUI-KJNodes for MiniMaxLowVRAMAttention and MiniMaxChunkFeedForward, and the media handling needs ffmpeg on PATH. You'll also need the H3 FL2VA base model, the Qwen3-VL text encoder and both VAEs where the README says.

The realistic failure modes

Out of memory is the obvious one: 320×192×39 runs fit in a 16GB card's spare VRAM, a 768-frame 736×608 job is a different animal. Lower generation_megapixels in the Plan, or shorten window_frames, before you lower steps. The second is cache confusion - change the source, geometry, prompt or VAE and you need a new run_name; change only seed, steps or model and you need a new candidate_name if you're on the candidate route. Getting that wrong produces identity-check errors that look like bugs and aren't.