MiniMax H3 Local MV Vocal Lock Renderer V2 / 独立人声锁定生成 (Advanced EXP/T8)

The V1 renderer conditions every scene on the full song mix, and that's its weakness: the model has to work out which part of a loud, busy master is the voice worth syncing to. This V2 renderer fixes the input problem at the source - it demands a separate, timeline-aligned isolated vocal track and feeds that into each H3 scene's lock_source, so the mouth is driven by clean vocals instead of an inference the model has to make over a wall of instruments.

The two-track contract

Concretely, you connect full_song and vocal_lock_audio. The full song never enters H3 and never conditions any scene - it's held back and muxed exactly once onto the finished, assembled video at the end. The isolated vocal (or clear dialogue) is what each scene's H3 lock_source window actually hears. Two roles, deliberately split: vocal for the face, master for the final render. This is the "vocal lock" architecture that separates this whole V2/V3 line from the compatibility V1 route.

Everything else about the renderer is the serial-resume machinery the V1 node introduced, so I won't re-litigate it: scenes run strictly one at a time through your connected local model (with the Qwen3-VL clip and the video/audio VAEs), accepted scenes are saved atomically under a contract bound to your chain_id (default my_h3_local_mv_vocal_lock_v2), and re-running the same chain after an interrupt resumes instead of restarting. No HTTP /prompt, no remote anything - fully local and serial by design.

Inputs and outputs

Inputs that matter:

• model, clip, video_vae, audio_vae, reference_image (performer identity), full_song, vocal_lock_audio, and the mv_vocal_lock_prompt_plan from the Vocal Lock Prompt Compiler V2.
• width/height - default 1056×608, multiples of 32.
• steps - default 8, with sampler_name/scheduler defaulting to dual_clock_euler/native_flow and shift_video/shift_audio at 6/3. Note the tension I'll flag below: this renderer defaults to the 8-step recipe that the pack's own A/B testing later abandoned for the official 4-step route.
• chain_id, base_seed, resume_existing (true), filename_prefix, crf/bit_depth for encode, model_id (audit-only).

Outputs: video + preview, video_path, manifest_path, completed_scenes, status, report_json.

V2 or V3? Read this first

Now the part that decides whether you use V2 or jump to V3. The pack's own history is blunt here: V2's single-scene clear-speech sample passed SyncNet (0-frame offset) and normal-speed human lip-sync review, but the performer's surroundings looked soft, and a later V2-era attempt died on a giant background duplicate face. Their controlled follow-up - same image, same audio, same failing seed - traced the culprit to the recipe bundle, not the seed or the H3 base model: they had combined a generic LarryVrh EMA Turbo LoRA with a non-official 8-step / shift-6:3 Ref2VA schedule. The accepted fix pins the official Ref2V Turbo v0.1 LoRA at 4 steps, Euler/simple, shift 12/3 - and that configuration is exactly what the V3 renderer ships as its defaults. So V2 remains in the pack as the historical 8-step Ref2VA route and for lighter single-clip work, but the README's recommendation for verified long-MV lip-sync is the V3 chain (same V2 planner, Visual Director V3 for the one-person/one-face shot contract, and the V3 renderer).

If you use V2 anyway - say, you want the 8-step quality headroom on a short clip - the lesson to take from their findings is: don't bolt a random EMA Turbo LoRA onto a Ref2VA schedule, and expect double-face ghosting if you do. Start locked-off on the camera too; dynamic moves are where subject edges smear.

Installing it

Install is pack-standard: ComfyUI Manager search MiniMax H3 Audio T8, or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git into custom_nodes, restart, and update ComfyUI core first. Source your isolated vocal locally before you start - this node won't separate stems for you, and the whole route fails if vocal_lock_audio isn't genuinely timeline-aligned with full_song. The dated V2 workflow in examples/workflows/24-mv-lipsync shows the correct dual-audio wiring.