MiniMax H3 MV Vocal Lock Prompt Compiler V2 / 官方六段式口型提示词 (Advanced EXP/T8)

The V1 prompt compiler gives you a solid per-scene Ref2VA prompt. This V2 compiler is what you move to when "solid" isn't good enough and you want the exact prompt structure the official MiniMax H3 Ref2VA format expects, built with the discipline that got the pack's 32-second five-scene MV through user acceptance. The difference is in the word official and in one behavior: lyrics and dialogue only ever enter the prompt as verbatim text you supplied, inside a <d> marker - never as a guess.

The official six sections

Official means the compiled prompt follows the six-section order MiniMax's Ref2VA format uses, verified in the pack's own validator: subject_definitions → summary → retention_analysis → detailed_description → overall_soundscape → non_diegetic_music, each appearing exactly once and in that order, or the prompt is rejected. If you've ever watched H3 silently ignore a hand-written wall of prose, this is why structured beats freeform here: the model's retention logic keys off <Subject 1>, <Picture 1> and <Audio 1> markers, and the compiler wires them deliberately rather than hoping you typed them right.

The identity and audio contract it encodes:

• <Subject 1> comes from <Picture 1> - the lead performer stays the person in the reference image, with fully_preserved marking identity/face/hair/clothing retention.
• <Audio 1> is the isolated vocal-lock track, reused as the synchronized vocal signal with a fully_copy marker - the clean vocal drives the mouth 1:1.
• Vocal scenes get forced framing: continuous medium close-up, front or three-quarter face, unobstructed mouth. Unsafe wide/rear/profile setups are rewritten rather than passed through.

Inputs and outputs

Inputs: scene_plan (from the Vocal Lock Scene Planner V2), then the same style family as V1 - global_creative_prompt, performer_description, visual_style - plus the V2-specific set:

• vocal_content_type - singing or spoken_dialogue; the compiler words the mouth contract differently for each.
• vocal_language (advanced, default English) - labels the <d> text.
• exact_vocal_text_json (advanced) - per-scene verbatim lyrics/dialogue. Only this text ever enters <d>; leave it empty and the prompt simply never puts words in the mouth.
• non_vocal_action (advanced) - what the performer does between phrases.
• camera_pattern - and here's a V2 gotcha the tooltip is upfront about: the default is a locked-off static camera with continuous mouth visibility. Moving the camera is allowed - push-in, handheld, lateral - but the compiler's warning is real: dynamic camera work raises temporal ghosting risk on the subject edges, which is the exact artifact that sank earlier versions of this route. Start locked-off.

Outputs mirror V1: mv_vocal_lock_prompt_plan (typed, feeds the V2 or V3 renderer), segment_prompts_json, prompt_relay_events, prompt_preview, report_json. Preview before you render a 20-minute serial job - the compiler is deterministic, so the preview is a truthful sample of what will run.

Why the pedantry pays off

Why the pack is this pedantic about it: their own A/B testing found that earlier failures - soft performer edges, double-face ghosting - were not the seed or the base model but a mismatched prompt/recipe bundle (a generic EMA Turbo LoRA on a non-official 8-step schedule). The official six-section structure plus the official Turbo recipe is the difference between a clip that ghosts and one that holds. That's the value of a compiler that follows the format exactly: when it fails, you know it's your material, not your prompt formatting.

Installing it

Install via ComfyUI Manager (search MiniMax H3 Audio T8) or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git into custom_nodes, then restart. Keep ComfyUI core itself updated - this pack's nodes need the recent native H3 APIs and go red en masse on older builds.