MiniMax H3 Local MV Vocal Lock Visual Renderer V3 / 单人物视觉合同生成 (Advanced EXP/T8)

This is the renderer at the end of the route the T8 pack actually vouches for. Where V2 defaulted to 8 steps with a generic Turbo bundle - the combination their own A/B testing proved caused the ghosting - the V3 renderer ships with the official recipe baked in as its defaults: 4 steps, Euler/simple, video shift 12, audio shift 3, the official Ref2V Turbo v0.1 LoRA at strength 1.0. And it only accepts prompt plans that carry the Visual Director V3's one-person/one-face contract. Same serial local generation, same resume, same "full song muxed once at the end" architecture as V2 - but the config that the author validated is the config you get by default.

The recipe is the defaults

Read the steps tooltip before you touch anything: official Ref2V Turbo v0.1 uses 4 NFE - don't run this distilled LoRA as an 8-step job. The distilled 4-step model isn't a 4-step approximation of an 8-step model; treat the step count as part of the recipe, like salt. Same for sampler_name/scheduler (euler/simple) and the shift pair (12/3). The pack's earlier V2/V3 r1–r3 failures were traced precisely to substituting a generic EMA Turbo LoRA and an 8-step/shift-6:3 schedule on Ref2VA - with the official bundle, the same failing seed stopped ghosting. V3 exists so you don't have to know that history to avoid repeating it.

Inputs and outputs

Inputs:

• model, clip, video_vae, audio_vae (local graph wiring), reference_image (a clear front or three-quarter-face identity shot - the tooltip asks for it), full_song, vocal_lock_audio, and mv_vocal_lock_prompt_plan from the Visual Director V3.
• width/height default 1056×608 (multiples of 32). The published recommended workflow pins 1024×768 - both are fine; pick one and keep the model's native geometry happy.
• steps 4, shift_video 12, shift_audio 3, euler/simple - the official recipe.
• chain_id default my_h3_local_mv_vocal_lock_v3 - resume handle. base_seed, resume_existing (true), crf/bit_depth, model_id (audit-only; its default value literally records the expected Ref2VA+Turbo4 combo).

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

Why this one over V2

Why people reach for this over V2: it's the configuration that produced the pack's accepted long-MV proof - a 32-second, five-scene master at 1024×768/24fps where all 5 scenes decoded clean (strict H.264/AAC, 20-repeat decode with zero anomalies), no duplicate faces or persistent edge ghosting in scene sampling, official SyncNet offsets of 0/-1/0/-1/0 across the five isolated-vocal scenes, and the user's verdict after full playback: the 32-second result was fine - perfect - which led them to cancel a further ~90-second request. That's about as strong an end-to-end datapoint as an Advanced EXP pack publishes, and it's the reason the V3 workflow is the recommended entry in 24-mv-lipsync.

Keep the honesty that the author keeps: the acceptance is bound to that master's SHA-256, manifest accepted means mechanically saved and contract-bound rather than human-approved on your new footage, and SyncNet proves temporal AV alignment within a frame - not universal per-phoneme linguistic correctness. New material still needs a normal-speed human watch for identity, hands, background and cuts.

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, restart, with ComfyUI core updated first (the pack needs the recent native-H3 and weight-adapter APIs). Put the official minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors LoRA in models/loras, use a Ref2VA base (not FL2VA - don't mix), and load the dated VocalLock_V3_Official_Ref2V_Turbo4 workflow rather than hand-assembling the chain. Also note the license reality once: these local H3 weights are excluded for US, EU, UK and South Korea users.