MiniMax H3 Prompt Relay Long Video Conditioning / 长视频分段路由条件 (Advanced)

Putting Prompt Relay into a Long Video pipeline is a composition problem: Long Video already does its own thing - re-injecting the previous segment's tail as motion context - and relay wants to patch the model's attention with timed event windows. Do them naively and the two mechanisms fight over the same attention entry point. MiniMaxH3PromptRelayLongVideoConditioningT8Advanced is the node that composes them: it takes the existing Long Video motion-context payload repair and the Prompt Relay attention routing, and isolates them so each does its job without clobbering the other. The description's phrase is "isolated composition" - and the promise is that the old Long Video Conditioning schema is never changed.

This is the per-segment workhorse. Where MiniMaxH3PromptRelayLongVideoPlanT8Advanced projects the global timeline into a segment's window, this node consumes that projected plan, encodes it through the H3 Qwen chain, and produces the same downstream bundle the plain relay Conditioning node does - model, positive, av_latent, mux_audio - but with Long Video's context handling layered in.

The inputs that matter

• model, clip, video_vae, audio_vae - native H3 components, unpatched (the tooltip on model says "not connected to other patches").
• context (H3_T8_CONTEXT) - the Long Video context handle, wired from the Planner.
• prompt_relay_plan - must come from the paired Long Video Window node (MiniMaxH3PromptRelayLongVideoPlanT8Advanced). Not the plain relay Plan - that's a hard requirement.
• segment_index and context_frames (0–39, the 5/22/39 overlap) - which segment this is and how much past it re-renders.
• length - 124 default, steps of 17, on H3's native frame grid.
• context_audio - video_and_audio by default; the audio side of the context tail is a real choice here.
• The familiar relay/audio bundle: task_type, audio_mode (native default), audio_denoise_strength (0.35), add_source_as_reference (true), execution_mode (report_only → apply_exp), query_chunk_rows (256).
• Long Video extras: first_frame_reuse (segment0_only), plus persistent_identity_image / persistent_identity_strategy / persistent_identity_interval - the identity-through-segments levers.

Outputs: model, positive, av_latent, mux_audio, conditioned_prompt, media_map_json, report_json.

The wiring rules that matter

The same discipline applies as the plain relay Conditioning node - report_only first, then apply_exp; lock_source + mux_audio to the save node if you need the original voice; and the README's reminder that drive_audio is a generation condition, not deterministic lip-sync. The real chain the author ran - segment 0→1 with 22-frame context, events that didn't restart at the head, 124+102 frames, three rounds of strict decode on video and audio - validates the mechanics, not universal quality. Audio seams still need your ears.

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; supply H3 weights and VAEs; ffmpeg on PATH for final Long Video concat. The template is examples/workflows/04-long-video/2026-08-20_H3_Prompt_Relay_Long_Video_Turbo8_Advanced_EXP.json - follow its canvas notes for the exact ordering.