MiniMax H3 Prompt Relay Query Route / 联合音画路由 (Advanced)

Prompt Relay, as published, routes video - local event prompts get attention windows on the target video time axis. H3 is a joint audio-video transformer, so the natural question is "can the same trick route the audio side too?" MiniMaxH3PromptRelayQueryRouteT8Advanced is the explicit opt-in for that, and the way it's built tells you exactly how much the author trusts the idea.

Default query_route is video_only_paper: the plan stays on the paper-faithful video route, which the README describes as "not changing old workflows." Only when you explicitly pick joint_av_exp does the node also route the target audio on H3's native audio time axis - and the tooltip is blunt that this is an unverified experimental extension, not a capability the paper validates. This is the pack's fail-closed ethos in one dropdown: the safe default is the published math; the exciting mode exists but wears a warning label.

What it does

It's a tiny, surgical node: input a prompt_relay_plan and a query_route, get the plan back plus a report_json. Nothing else. You insert it between the Plan and the Conditioning node - the official wiring is Plan → Query Route → Conditioning. If it's not connected at all, the old workflow bypasses it entirely and the Conditioning node routes video-only, same as always. That "append-only, old workflows unchanged" pattern is the pack's signature move.

Two things worth knowing before you flip to joint_av_exp. First, from the Conditioning node's own tooltip: joint AV routing "cannot be combined with lock_source" - you can't route the target audio and insist on keeping the source audio unchanged; those goals fight each other, and the node refuses the combination rather than silently picking one. Second, the mechanics: H3's audio runs on its own 40Hz clock, separate from the 24fps video clock, so routing the audio side means aligning event windows to that second time axis - which is exactly what makes it a real implementation instead of a copy-paste of the video math.

Where it fits

Use it when you're deliberately exploring relay-with-sound - an event like "the door slams at second three" that should influence the audio generation at that moment, not just the picture. The pack's Joint AV template (2026-08-20_H3_Prompt_Relay_Joint_AV_Turbo8_Advanced_EXP.json) shows it wired for real. Just keep the README's honest caveat in your head: the joint-AV path is an H3 experimental extension, and the author's own mechanical A/B tests prove the timeline mechanics - not that routed audio always sounds better or is perceptually indistinguishable.

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, no weights required to route a plan. Start on video_only_paper, read the report_json, and only reach for joint_av_exp when you know what you're buying.