MiniMax H3 Audio Refine Long Video Delivery Split / 长视频精修交付分流 (T8 Advanced EXP)

Long video with MiniMax H3 is a game of latents. Each segment gets generated, its joint AV latent is kept as context for the next segment, and if you're using the pack's Audio Refine chain you now have two versions of that latent floating around: the original segment latent, and the reviewed one whose audio slice you just refined. This small node exists to make sure those two never get mixed up - because feeding the wrong one back as continuation context quietly poisons everything downstream.

Here's the scenario in plain terms. You generate segment N, you like the picture but the audio is off, so you run the audio refine chain and get a candidate. You listen, you accept it. Now you have a latent that's the "good delivered" version of segment N - but if you hand that to the context node as the starting state for segment N+1, the next segment inherits a state that isn't the clean continuation the planner expected. The author's warning is blunt: never feed delivery_av_latent into the next segment.

What the node does

It takes both latents and splits them cleanly:

• original_continuation_av_latent - the exact, untouched segment latent. This is the one that flows into the next segment's context. Unmodified, by design.
• reviewed_delivery_av_latent - the latent after your review/refine. Out of this, the node emits a delivery latent whose video is forcibly restored from the original, so the refinement can never silently change the picture - only the audio carries the refine.
• candidate_selected - the boolean that decides whether the refined audio is actually used in the delivery output. Default is false, matching the pack's philosophy: default to the original, accept the candidate only when you've listened and chosen it.
• segment_index - bookkeeping so the report can tell you which segment you're splitting. Handy when you're batching a whole timeline.

Outputs are the two separated latents - continuation_av_latent and delivery_av_latent - plus a report_json. The continuation output plugs into your long-video context node; the delivery output heads to decode and save.

Why this matters more than it looks

The pack's long-video verification shows this is not paranoia. They ran a real test where a refined candidate's video latent, if left unguarded, would drift from the original segment's picture - which is exactly the kind of subtle seam that destroys a multi-segment shot. And the audio-refine README is blunt about what can change in a refine: wording, added or dropped syllables, the character of a voice, performance, music, SFX, ambience, even lip-sync. If you accept a candidate and then reuse that latent as context, you're baking a different "world state" into the next segment. This node is the guardrail that keeps the two lanes separate.

Installing and where it lives

It's part of the same pack as all the other Audio Refine nodes, so install once:

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

or ComfyUI Manager → search "MiniMax H3 Audio T8" → restart. Update ComfyUI itself first or the whole pack import-fails; the requirements.txt is intentionally empty because ComfyUI supplies the stack.

Start from the 18-audio-refine long-video workflow rather than hand-wiring this - the README warns the long-video + Prompt Relay + Turbo8 combo is a "run separately, don't queue all at once" kind of workflow. And keep the discipline: the continuation latent is sacred. If you ever catch yourself wondering why segment N+1 looks subtly wrong, check which latent you fed the context node.