T8 Video Prompt Compiler / 多后端提示词编译 (Advanced)

The name is a little cheeky: "T8 Video Prompt Compiler" isn't really a video compiler at all. It's a text compiler - one structured visual-and-audio brief that can target MiniMax H3, Wan 2.2, LTX-Video, or a generic cinematic backend from the same inputs. The node description is blunt about it: "This is text compilation only." No API calls, no keys, no model loading. It takes your prompt, optional dialogue, cast, sound canvas, and creative brief, and produces a tidy prompt_packet plus separate compiled positive, negative, and audio prompt strings.

Why would you want that? Because H3 is a joint audio-video model - your prompt isn't just describing pictures, it's describing a soundtrack, ambience, and dialogue in the same breath. Most people write a visual prompt, get an audio prompt wrong, and wonder why the sound is a coin flip. This node splits the job: the visual brief and the audio brief get compiled together from one structured source so they actually agree.

The inputs that matter

• prompt - the raw brief. Everything else is structure around it.
• backend - minimax_h3 (default), or Wan 2.2 / LTX-Video / generic cinematic. Same brief, different target syntax.
• duration_seconds (5.167 default) and aspect_ratio (16:9) - the canvas facts the compiler bakes into the output.
• strict_exact_dialogue - on by default; keeps spoken dialogue verbatim instead of letting the compiler paraphrase it. If you have a line of dialogue the character must say, this is the flag that protects it.
• dialogue, negative_prompt - the obvious companions.
• cast (H3_T8_UNIFIED_CAST) and sound_canvas (H3_T8_SOUND_CANVAS) - the pack's Studio data types. These are the interesting wires: the README describes "Unified Cast + Sound Canvas + Prompt Compiler" as the global fact source for picture and sound in a Studio workflow. The relay nodes downstream won't auto-split your story or rewrite events - the compiler is where the authoritative global description gets built.
• cast_ids, creative_brief_json - optional structured inputs for more elaborate setups.

The outputs that matter

• prompt_packet - the H3_T8_PROMPT_PACKET handle that feeds the Packet → Relay Plan bridge (MiniMaxH3PromptPacketRelayPlanT8Advanced). That's the handoff: compile once, then build a relay timeline on top of it.
• compiled_prompt and negative_prompt - ready to wire into your conditioning chain.
• audio_prompt - the compiled audio-side brief, which you'd otherwise have to write by hand.
• report_json - what went in and how it came out.

Where it fits and what to watch

If you're comparing H3 against Wan 2.2 on the same concept, this is genuinely handy: swap the backend and the same brief recompiles for a different model family, which makes A/B comparisons far less painful than maintaining two divergent prompts. Just remember the honest limits - it compiles text; it doesn't guarantee H3's conditioning will follow it, and the audio prompt is only as good as the source brief.

Install

Ships in 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. The Studio examples live in examples/workflows/11-studio-production/, and the relay handoff example is examples/workflows/14-prompt-relay/2026-08-20_H3_Studio_Prompt_Packet_Relay_Stock20_Advanced_EXP.json.