MiniMax H3 Prompt Relay Long Video Window / 长视频绝对时间窗口 (Advanced)

Long Video generation in ComfyUI works by rendering overlapping segments: each new segment re-renders a context window of the previous one, and you stitch them back together. That's fine for plain prompts. It's a disaster for Prompt Relay, because if you naively re-apply the same timeline to every segment, your local events restart from the beginning at each segment boundary - the hand-raising that should be a one-time beat happens over and over, and timed audio events land on the wrong frames.

MiniMaxH3PromptRelayLongVideoPlanT8Advanced is the fix: it projects one global relay timeline into the exact rendered window of a specific Long Video segment. The key math is in the description - continuation frame 0 maps to accepted-start minus context overlap - which is a fancy way of saying the plan gets re-based so events keep their position on the accepted timeline even though the segment physically starts earlier (inside the context overlap).

The inputs that matter

• prompt_relay_plan - the global plan, created once for the whole video. You don't build a new plan per segment; you project the same one.
• segment_index - which segment this window belongs to.
• length and context_frames - wire these straight from the Long Video Planner, per the tooltips. context_frames is the 5/22/39 overlap; this node needs to know exactly how far the segment re-renders into the past.
• timeline_start_seconds / timeline_end_seconds - the accepted-timeline window this segment covers, in absolute time.

Outputs: the re-based prompt_relay_plan (consumed by the Long Video Conditioning node), compiled_prompt, and report_json.

How it behaves in practice

The author ran a real chain: 736×416, Turbo8, 22-frame context, segments 0→1, and the relay events did not restart from the head - output was 124+102 frames with video and audio each passing three rounds of strict decode, at a whole-card peak around 15478/14984 MiB. Read that honestly: it proves the timeline mechanics and the continuation math. The README's own conclusion is that final audio seams still need listening, and none of this is a general quality or 16GB-safety claim. What you get is certainty that your events sit where you put them across segment boundaries - which is the entire point of the node.

Where it fits

Chain it with MiniMaxH3PromptRelayLongVideoConditioningT8Advanced: the plan projects the window, the conditioning composes it with the Long Video motion-context repair and installs the relay patch. The README's Long Video + relay example is examples/workflows/04-long-video/2026-08-20_H3_Prompt_Relay_Long_Video_Turbo8_Advanced_EXP.json. And remember the general rule that also applies here: the global Plan is created once; per-segment, you only re-project, never re-invent.

Install

Part of the MiniMax H3 Audio T8 pack:

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

or search "MiniMax H3 Audio T8" in ComfyUI Manager and restart. No pip deps; you supply the H3 weights. Long Video final concat additionally wants ffmpeg on PATH - most ComfyUI builds have it, and the node errors clearly before writing if it's missing.