MiniMax H3 • Split AV Streams

The thing that makes MiniMax H3 different from every video model you've used before is that it doesn't generate a silent movie and bolt audio on afterwards. It treats picture and sound as one joint latent - a NestedTensor that packs a 24-channel video stream and a 32-channel stereo audio stream together. That's great for lip-sync and sound design, and awkward for editing, because almost every tool you know wants to touch one or the other. MiniMaxH3LatentLabSplitAV is the seam you pull to get at the two halves.

Give it an av_latent (LATENT) and it returns two LATENTs: video_latent and audio_latent. Nothing else - no settings, no knobs. The video half is the [B, 24, T, H, W] H3 picture stream; the audio half is H3's 40 Hz audio latent carrying 32 channels of stereo. Split them, edit whichever side you care about, then pack them back together (the pack has a Pack AV node for that) and hand the result to the sampler or decoder.

Why you'd actually use it

This node is the foundation of every "keep the sound, change the picture" or "keep the picture, change the sound" trick in the pack. Concretely:

• Lip-sync setups - split the AV latent, replace the audio stream with your performance source, re-pack, and let the sampler regenerate only what it needs.
• Reference/editing work - extract the video stream to feed a video_ref_edit or inspection path while leaving audio untouched.
• Stream-level control - once split, you can route the video half through an encoder or the audio half through its own denoise control, then recombine.

If you're doing manual A/B tests, splitting also lets you validate one stream's geometry without the other getting in the way.

Install

It's one node in one pack, and the pack installs with no extra Python dependencies:

cd ComfyUI/custom_nodes
git clone https://github.com/vizart-vj/ComfyUI-MiniMax-H3-LongMedia

Restart ComfyUI, or use ComfyUI Manager and search "MiniMax H3 LongMedia". The pack itself is code-only - the MiniMax H3 model weights, H3 video VAE and audio VAE belong in ComfyUI's regular model folders via its native H3 support. You also need an actual joint AV latent to split, so this node only appears once you're feeding it from a Setup or an encoder output.

Gotchas

• The two halves only make sense together again if their durations match when you re-pack. If you trim one stream, expect the pack to complain - H3 audio and video latents have a fixed relationship (24 fps video, 40 Hz audio latent).
• This is a Streams-category utility, not something you tune. If you find yourself reaching for it constantly, the pack's per-stream denoise and replace nodes were built to save you the round-trip.
• Same pack-wide caveat that applies to all of this: the H3 open weights are excluded from the US, EU, UK and South Korea under the MiniMax H3 Community License.

One input, two outputs, zero settings. It's the most boring node in the pack and one of the most useful - every interesting AV workflow in LongMedia starts by pulling this seam.