H3 Video Outpaint · 接续所选候选 (EXP)

The candidate node gave you window one. MiniMaxH3VideoOutpaintContinueCandidateT8 does everything else - it takes the confirmed selected handle, walks the remaining windows serially inside one execution, commits each to the run's cache, and hands back a sampled handle that the Compose Candidate node turns into a file.

Two inputs, and that's genuinely all: model and selected. No seed field, no step count, no mode switch. That's the design. The settings come from the selection archive, which is the same record containing the first frame you approved. You can't accidentally speed it up, slow it down, or flip the decode mode halfway through a video.

How the continuation actually works

The plan already divided your clip into shots (based on cut_frames_json in the Plan node) and each shot into windows of window_frames frames. Window one is done. This node generates the rest, in order, reusing the prepared source, audio and prompt caches from Prepare, so nothing gets re-encoded. The prepared caches are re-validated against the plan's source hash before anything runs - swap the source file underneath a run and you get an error rather than a video that's half one clip and half another.

Because windows are committed as individual assets rather than held as one big batch, a crash costs you the current window and no more. Restore the selection by ID, wire it back in with model, and the committed windows get reused.

There's a subtlety in the plan geometry worth knowing even though this node doesn't expose it: the sampling canvas is aligned to H3's latent grid using available slack, so a partial generated strip never gets locked in as source context. That alignment is why the delivered geometry can be odd while the sampling geometry stays on-grid, and why the resulting frame doesn't look like the source got nudged a pixel sideways.

The audio

Your original soundtrack rides along in the prepared audio cache and is carried to delivery. This node doesn't generate new audio, and the pack is explicit that a silent source stays silent rather than being silently filled with music or dialogue. If the source had an audio stream and you didn't connect the audio VAE at Prepare, you'd already have been stopped there.

Outputs

sampled - feed it to MiniMaxH3VideoOutpaintComposeCandidateT8 to get the MP4, or save the ID if the job is going to span more than one sitting. sampling_report is JSON describing what got committed.

Realistic expectations on time and VRAM

The pack's reference job was 768 frames at 736×608 as 22 serial windows. That is not a coffee break; on a 16GB card it's an evening, and the docs' advice is one H3 job at a time - don't run a second generation alongside it. Lower generation_megapixels in the Plan before you consider lowering steps. And note that resume behaviour isn't exposed here; you recover by loading the selection or the completed cache and re-running, not by toggling a flag.

Install

Manager search: MiniMax H3 Audio T8, then restart ComfyUI fully. Manual:

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

requirements.txt installs nothing on purpose, so the pack can't replace your Torch/CUDA stack. This is a generation node, so it does need the real kit: H3 FL2VA base model in models/diffusion_models, the Qwen3-VL encoder in models/text_encoders, both VAEs in models/vae, ComfyUI-KJNodes for the audited low-VRAM attention and FFN patches, and ffmpeg on PATH. Do not stack Turbo, SLA, VDN, SPEED or Fast H3 on top - the compatibility matrix lists all of them as unsupported for this route, and "unsupported" here means the combination produces conflicts or wrong output rather than a clean error.

If it dies mid-run

Check whether the pack's nodes changed version. The caches verify identity across source, model, ComfyUI core, KJNodes and the sampling implementation, and an upgrade can legitimately invalidate an old cache - in which case you restart the run rather than fighting the check. The docs are blunt that you shouldn't try to edit the identity record to get past it. Also possible: a Windows crash kills the process outright. Recover from the last completed cache and move on.