MiniMax H3 Prompt Relay Conditioning / 分段提示词条件 (Advanced)

Everything before this node is just paperwork. The Plan holds your global prompt and timed events; MiniMaxH3PromptRelayConditioningT8Advanced is where the relay becomes real: it authoritatively encodes the plan through the native H3 Qwen chain, binds exact local token spans, and optionally patches a cloned H3 MODEL with a bounded-memory, plan-selected query relay. Output is a new model, positive conditioning, and a ready av_latent - the whole downstream chain, in one node.

If that sounds like a lot of moving parts, it is, and the author has engineered it so you can't blow yourself up by accident. Default execution_mode is report_only: it encodes and reports token/layout/VRAM numbers but does not install the attention routing. You check the report, then explicitly flip to apply_exp. The node also self-limits: with zero local events it passes through unchanged, and with exactly one event it installs no patch either - a single event belongs in global_prompt, not a relay.

The inputs that matter

• model, clip, video_vae, audio_vae - native H3 components. The model output is a local clone, so your original MODEL stays untouched elsewhere in the graph.
• prompt_relay_plan - must come directly from a same-version Plan node. Plan mismatch is a hard error, not a warning.
• task_type - T2VA/I2VA/FL2VA/L2VA/Ref2VA/Hybrid. Relay routes the target video time axis for all of them; extending to target audio requires the separate Query Route node with joint_av_exp.
• audio_mode - native, lock_source, remix, or reference. Want the original audio preserved? Use lock_source and wire the node's mux_audio output to your save node - not AV Decode's generated_audio. The README is adamant about this wiring.
• audio_denoise_strength (0.35), add_source_as_reference (true), prompt_primary_audio_ordinal, strict_prompt_tags (true) - the audio-conditioning bundle.
• execution_mode - report_only → apply_exp. Start here, always.
• query_chunk_rows (256) - smaller chunks save VRAM but run slower; the report tells you if you're near a cliff.

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

Ordering rules that will bite you

With Turbo8 the README requires the chain UNET → Prompt Relay Conditioning → Alpha8 Bypass LoRA → DualClock Sampler - put the LoRA before the relay and it's actively rejected. And drive_audio is a joint generation condition, not a deterministic lip-sync guarantee: if you need the exact input voice, lock_source + a <Audio N> reference is the strongest option, but don't expect per-phoneme mouth sync without a dedicated face/lip tool afterwards.

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 H3 weights, Qwen3-VL CLIP, video and audio VAEs yourself. The eleven relay templates under examples/workflows/14-prompt-relay/ each wire this node, with per-task wiring notes on the canvas - start with the Stock20 T2VA template and read the report_only output before you ever switch to apply_exp.