MiniMax H3 Masking · Master Audio + Video Prefix

MiniMax H3 generates audio jointly with the picture, which is great until you already have the music or dialogue and need it to survive the render note-for-note. This node is how you get that: it drops an exact slice of your master audio into H3's complete audio target, then masks that interval so denoising can't touch it. If you've watched H3 "interpret" a song and drift into a slightly different take, this is the node that stops the improvisation.

What it does

The full master_audio timeline (music, dialogue, narration, effects - anything you already have) is connected once. The node reads clip_start_seconds to find the current raw H3 clip's interval, VAE-encodes that exact slice into the target latent via the audio_vae, and protects it with an all-zero audio denoise mask. The target's audio is authoritative - the reference bank and any carried predecessor can't compete with it. That's the same "lock source audio" behavior the Chain Policy switch enables, but as a discrete node that works on your prerecorded master rather than the source timeline.

The video side is optional: context_length (native runs like 5, 22, 39, 56…) requests a protected previous-video prefix. Two ways in:

• source_latent - the preferred live path. It copies the phase-aligned tail of the previous sampled H3 video latent directly. Its audio is ignored, because master_audio is authoritative.
• source_frames - the legacy decoded path. Requires the video vae, plus source_fps and crop (disabled stretches, center preserves aspect and crops). Use this only when you don't have a latent to hand off.

The three outputs you actually care about

• latent - the sampler target, with exact protected audio and the optional video prefix baked in.
• trim_frames - how many decoded frames the protected visual prefix occupied. Feed it to Loop Trim so the repeated head comes off picture and sound together.
• clip_audio - the exact master-audio interval this raw clip represents, handy for review or a downstream mix.

Install and gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/ethanfel/ComfyUI-MiniMaxH3-Contex-Loop.git

Restart, or grab it from ComfyUI Manager under "MiniMax H3 Contex Loop". No pip step; you do need a current ComfyUI (native Add Guide for MiniMax H3 from PR #15439) and ideally ffmpeg on PATH. Models and VAEs aren't bundled - you bring the H3 weights plus video and audio VAEs, and remember the MiniMax H3 Community License restricts local weights outside the US/EU/UK/Korea block.

The common mistake is feeding the whole movie into Ref2VA and treating it as the audio master. Keep one job per wire: this node owns the exact target clock, and your source video shouldn't double as a loose reference. And don't zero out context_length expecting to keep the previous clip's audio - with master audio connected, the master is the boss; context_length only controls the visual prefix.