MiniMax H3 MV Vocal Lock Visual Director V3 / 单人物逐镜头导演 (Advanced EXP/T8)

If you've generated reference-based H3 video of a person, you've met its favorite party trick: it slides a second, slightly-off copy of your subject's face into the frame - a background face, a mirror, a poster, a projection - and ruins the shot. The pack's own history is a graveyard of those failures (a V2 run died on a giant background duplicate face; V3 iterations ghosted until the recipe was pinned). This node is the attempt to write that problem out of the prompt by making the shot itself a legally binding document: one person, one face, nothing that could plausibly be a second one.

A director with a one-person contract

The Vocal Lock Visual Director V3 is the V3 successor to the V2 Prompt Compiler. Same plumbing - it takes a scene_plan from the Vocal Lock Scene Planner V2 and compiles one explicit six-section Ref2VA prompt per scene - but where V2 gave you a solid official-format prompt, V3 adds per-scene shot direction and an enforced visual contract:

• Exactly one performer and one human face in the entire frame.
• Explicitly forbidden: mirrors, reflections, projections, screens, posters, portraits, ghosting, duplicate faces, double exposure, background people and visible props.
• Each scene gets its own camera, lighting, performance and emotion direction.

A contract, not a suggestion

Why a "contract" rather than a suggestion: the compiler rejects scene direction that conflicts with it. You can't accidentally describe a shot that pans past a crowd or catches the singer in a reflection - the node refuses the text instead of hoping H3 ignores it. When your one-person render still grows a second face, the prompt structure is no longer the suspect; you can blame the material and move on.

Inputs you'll actually touch

Inputs you'll actually touch:

• scene_plan - from the V2 planner.
• global_creative_prompt, performer_description, visual_style - the recurring identity/style text.
• scene_directions_json (optional) - per-scene direction objects (camera, lighting, performance, emotion), one per scene in order. Leave it empty and the node applies a safe studio arc - locked framing, consistent lighting - which is a perfectly sane default for keeping identity stable across cuts.
• vocal_content_type - singing or spoken_dialogue.
• exact_vocal_text_json (advanced) - verbatim lyrics/dialogue per scene; only this text enters the <d> marker, and absent it the compiler never puts words in the mouth.
• non_vocal_action, vocal_language (advanced) - same roles as the V2 compiler.

Outputs are the standard compiler set: mv_vocal_lock_prompt_plan (typed, feeds the V3 renderer), segment_prompts_json, prompt_relay_events, prompt_preview, and report_json. Check prompt_preview before queueing a long serial render - it's the last cheap look you get before the GPU spends half an hour per scene.

The V3 chain it feeds

This node pairs specifically with the Local MV Vocal Lock Visual Renderer V3, which only accepts V3 visual-director prompt plans - it validates the contract on ingest and refuses to run a plan that doesn't carry it. So the chain is: Vocal Lock Scene Planner V2 → Visual Director V3 → V3 Renderer, running the official Ref2V Turbo v0.1 four-step recipe. That combination is what produced the accepted 32-second, five-scene master the README reports, with 5/5 scenes decoding clean and no duplicate faces in per-scene sampling. It's still marked Advanced EXP, and the docs are careful to say acceptance was bound to that specific master's SHA - a clean contract is a strong prior, not a guarantee on new material.

V2 remains in the pack unchanged; this is append-only. Choose V3 when you're doing a multi-scene lip-sync MV where a wandering second face costs you an hour of re-renders.

Installing it

Install: ComfyUI Manager search MiniMax H3 Audio T8, or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git into custom_nodes and restart, with ComfyUI core updated first. Grab the dated VocalLock_V3_Official_Ref2V_Turbo4 workflow under examples/workflows/24-mv-lipsync rather than hand-wiring the graph.