H3 V2V Init (nested AV latent)

H3 V2V Init is the middle of the MAINodes de-rope pipeline, and it does a job that doesn't look hard until you try it: it takes a VAE-encoded video latent and wraps it as the nested audio+video latent H3's SamplerCustomAdvanced actually expects. MiniMax-H3 denoises video and audio as one joint latent, so you can't just hand it a video tensor and go. This node builds the container, and then it exposes the dials that decide how much of the original survives the regeneration.

The core job

Input samples is the VAEEncode of your smeared frames (from H3 Time Smear). length defaults to 0, meaning "derive the frame count from the latent" - leave it at 0 and only set it to assert a specific grid. The output LATENT goes straight into H3's SamplerCustomAdvanced with a partial-denoise schedule (that's H3 Inject Schedule's job). The audio starts empty and generates jointly with the video on the truncated schedule - causally synced foley, which is the preferred audio source for regenerated content.

The dials a beginner actually sets

• oracle_samples + freeze_threshold - background freeze. Wire the baseline latent in, set freeze_threshold above 0 (0.35 is a sane start), and everything outside the oracle's motion region stays frozen to the smeared init. That's the fix for background agents - birds, crowds, traffic - that speed up through a held span: frozen background is held baseline content, so after exact recovery its timing is exactly the baseline's. freeze_grow (default 2) dilates the mask if effects fly too close to the subject and get clipped.
• mask + invert_mask + mask_feather - manual freeze, overriding the oracle path. The mask marks the region to regenerate; flip invert_mask to paint the background/birds to freeze directly. Default is hard latent cells (every ~16 px cell fully frozen or fully live, decode smooths the edge); raise mask_feather only if a hard seam ever shows. time_varying (off by default) keeps a mask's time axis so the freeze region can move - but a moving boundary can pop, so keep transitions on 17-frame phase.
• audio_latent + audio_mode - the dialogue savior. By default the audio rows start from zeros, so pass 2 invents a fresh performance at natural rate - and recovery then compresses those lips by the hold factor, so held regions come back rushed. That's the whole dialogue defect. Wire audio_latent from H3 Audio Smear → VAEEncodeAudio and set audio_mode to "follow the original performance (0.5)" and pass 2 renders a genuinely slowed take. Start there; "pin the original outright (0.0)" and the loose 0.7 exist for when the 0.5 bargain isn't right. audio_prefix_ticks lets you freeze just the first n ticks of the seeded audio so a continuation carries the previous segment's sound.

Install

Part of matlowai/ComfyUI-MAINodes:

cd ComfyUI/custom_nodes
git clone https://github.com/matlowai/ComfyUI-MAINodes

Restart, under latent/minimax/motion. No extra Python deps for the nodes; example graphs want ComfyUI-KJNodes, and you need MiniMax-H3 weights plus ComfyUI's H3 support (check the licence's US/EU/UK/Korea carve-outs).

The trap everyone hits eventually: the oracle freeze and the manual mask are mutually exclusive paths, and mask overrides the oracle. If you wire a mask and a freeze_threshold and wonder why the threshold did nothing - that's why. Pick one boundary-authoring tool per graph.