H3 Video Outpaint · 确认所连候选 (EXP)

Most ComfyUI workflows have no human checkpoint. You queue the graph, everything runs, you get a result. That's fine when a run is thirty seconds. It's a poor deal when the run is 22 serial windows of H3 video, which is why this pack builds the review in as a node rather than as advice in a README.

MiniMaxH3VideoOutpaintSelectCandidateT8 is the gate. Two inputs: candidate (the handle from the Candidate node, or from the Load Candidate node if you're resuming) and confirm_selection, a boolean that defaults to false.

Flipping it to true is you saying "I looked at this first frame and I want the rest of the video to look like this". Leave it false and the node throws rather than passing something through:

请先单独运行候选节点查看首帧，再开启 confirm_selection 确认所连接的候选。

Roughly: run the candidate node on its own and look at the first frame before you confirm. The check is is not True, so passing an expression that evaluates to a falsy value doesn't sneak past it either.

Why a node and not a prompt in the docs

Because the pack's whole design is that the settings you reviewed are the settings that continue. When you confirm, the node doesn't just bless the frame - it snapshots the candidate's settings and preview record into a selection archive, then hands back a selected handle. Every later stage reads from that archive. The continuation node uses the frozen settings; the composing node re-derives colour and geometry settings from the preview contract and can refuse to publish if the recomposed first frame's hash doesn't match what you approved. There's no path where you eyeball one thing and receive another.

The node also declines to be clever. Its description says outright that it won't auto-pick a candidate for you. That's aimed at a real failure mode in long-video tooling: automatic seeding and automatic acceptance produce beautifully consistent results that are consistently wrong.

Outputs

selected - wire it into the Continue node, or into the Compose Candidate node if the sampling is already finished. selection_report, a JSON string containing the selection plus the approved preview's hash. And selection_id, another full 64-character identifier. Copy the whole thing, verbatim. As with candidate_id, it's not a path and truncating it just gets you a lookup failure later.

Keep it. That ID is what lets you come back tomorrow, load the confirmed selection by ID and continue without re-sampling the first window - which on a 16GB card is real money in the electricity-and-patience sense.

The mistake to avoid

Selecting a candidate you generated with different settings than the ones you now want. If you've changed the seed in the Candidate node since, generate a new candidate with a new candidate_name and confirm that one. The archive binds settings, mode and geometry, and the checks downstream of here are strict enough that a mismatch surfaces as an error - usually well after you'd have preferred to know.

Doing this properly is cheap: the first window already exists, the review takes a minute, and the alternative is discovering at window nineteen that the join looked better in your head.

Install

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

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

No pip packages: the pack's requirements.txt is empty by design so installing it can't replace ComfyUI's Torch/CUDA stack. This node loads no model and touches no media - it validates and archives - so no H3 weights, no ffmpeg, no KJNodes needed to run it. That also means it fails fast and cheaply, which is the point: if confirm_selection is off, or the preview contract doesn't validate, you find out in a second rather than at the end of a queue.