MiniMax H3 Prompt Relay Plan / 分段提示词时间线 (Advanced)

Here's the problem every H3 prompt writer hits: you want the woman to look at the camera at second two, walk into the street at second four, and have the camera pull back at second six - but if you jam all of that into one prompt, the model smears every action across the whole clip. Nothing happens at the right time, and the scene feels like a shrug. The Prompt Relay technique fixes that by splitting the prompt into a global layer (scene, character, style, sound - always visible) and local layers (each action routed to its own time window). MiniMaxH3PromptRelayPlanT8Advanced is where you write that split.

The relay approach isn't T8mars's invention - the community knows it from GordonChen19/Prompt-Relay, which people run on LTX and Wan to get timed actions that actually land. What's new here is a faithful implementation for H3's joint audio-video transformer, with the math pinned to the published paper equation. The math_profile: paper_v1 default uses the epsilon/window/sigma equation from the paper; legacy_repo_compat exists only to reproduce old community parameter sets and is explicitly not the paper implementation. Use paper_v1.

The inputs that matter

• global_prompt - the always-visible scene/character/style/sound constraint. The default is a solid template: "Cinematic continuous shot, coherent subject identity, natural motion, stable lighting and synchronized ambience."
• local_prompts - one local action per line (or | separated). Each becomes an event. Clear text and time_ranges together and you get a global-only passthrough plan with no patch.
• length - 124 default, in frames; the node rounds up to H3's 17n+5 grid.
• timing_mode - auto_equal divides the timeline evenly; otherwise read time_ranges.
• time_ranges - one range per line (0-40\n41-81\n82-123), sorted by start time. Frames mode takes only integers with an inclusive end; seconds/percent treat the end as a boundary; percent is 0–100.
• epsilon (0.1), allow_gaps (false), allow_overlaps (false) - with gaps and overlaps off, your events must tile the whole timeline without doubling up. The author's rule is blunt: zero events = global passthrough, and a single event usually shouldn't install a patch either - fold it into global_prompt.
• prompt_relay_events - optional; a chained Prompt Relay Event node list becomes the authoritative event source.

Outputs: prompt_relay_plan (the handle everything downstream consumes), compiled_prompt, length, timeline_json, and report_json.

Where it goes

The plan is data, not execution - wiring it into MiniMaxH3PromptRelayConditioningT8Advanced is what actually encodes it through the H3 Qwen chain and patches a cloned model with the attention routing. You can preview it first with MiniMaxH3PromptRelayPreviewT8Advanced (which validates the timeline without loading any weights), or scope it with the Resource Estimate node. One honest note from the README: relay is a generation technique, so the pack's own mechanical A/B tests prove the timeline mechanics - not that relay always makes motion, identity, or audio better. It's a tool that makes timed actions possible; whether the result looks good is still on you and the model.

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 needed to build a plan - the relay example workflows under examples/workflows/14-prompt-relay/ each inline a Plan, and there's a pure plan/preview workflow (2026-08-20_H3_Prompt_Relay_Plan_Preview_Advanced_EXP.json) if you want to see a timeline before touching a model.