MiniMax H3 Prompt Packet → Relay Plan / 编译提示词转时间线 (Advanced)

The MiniMax H3 Prompt Relay system has two entry points: you can hand-write a plan with a global prompt plus local events (MiniMaxH3PromptRelayPlanT8Advanced), or you can compile a Studio brief in the T8 Video Prompt Compiler and bridge it over. This node is the bridge. It takes an authenticated prompt_packet from the compiler and converts it into an auditable relay timeline - while keeping the compiler's compiled visual/audio prompt as the global layer, untouched.

The key design rule, straight from the description: local events are never inferred or rewritten. The compiler produces a global fact source - the coherent picture-and-sound description that should hold for the whole clip. This node does not look at that global prompt and try to guess what local actions should happen; you supply those explicitly, either as an events_json list or by chaining Prompt Relay Event nodes.

The inputs that matter

• prompt_packet - the H3_T8_PROMPT_PACKET from the T8 Video Prompt Compiler. It only accepts backend=minimax_h3 packets with a complete hash. Compile for Wan 2.2 and try to relay it? Rejected. That's intentional - the relay math is H3-specific.
• events_json - the explicit event list. The default is a three-beat example ("The subject raises one hand" → "turns and walks forward" → "camera pulls back"). Two rules matter: in auto_equal mode you write only prompt (the node splits the timeline evenly); other modes need explicit start/end. And [] means "keep only the global packet, install no relay patch."
• timing_mode - auto_equal (default) or explicit frame/seconds/percent interpretation.
• math_profile - paper_v1 by default, matching the published Prompt Relay equation; legacy_repo_compat exists only to reproduce old community parameters.
• epsilon (0.1), allow_gaps (false), allow_overlaps (false) - the timeline-continuity controls. Gaps or overlaps off means events must tile the whole timeline.
• prompt_relay_events - optional chain of Event nodes; when connected, it becomes the authoritative event source over events_json.

Outputs: prompt_relay_plan, compiled_prompt, length, timeline_json, and report_json.

Why it matters

Prompt Relay is a real technique in this corner of the ecosystem (the community knows it from GordonChen19/Prompt-Relay on LTX and Wan): a global prompt always visible, plus local event prompts each routed to their own time window in the attention computation, instead of one monolithic prompt smeared across the whole clip. That's how you get "she looks at the camera and raises her hand" to happen at second two without the phrase leaking influence over the entire render. T8mars re-implements it for H3's joint AV transformer, and this node is how the Studio side of the pack talks to the relay side. The packet's compiled prompt stays global; your events stay yours.

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, then restart. No pip deps, no weights needed to build a plan. The handoff example is examples/workflows/14-prompt-relay/2026-08-20_H3_Studio_Prompt_Packet_Relay_Stock20_Advanced_EXP.json - wire the compiler's packet in, write your events, and read timeline_json before you ever load a model. If the node rejects your packet, check that you compiled for minimax_h3 and that you're connecting the compiler's prompt_packet output directly.