H3 Separate AV Latent

MiniMax H3 packs audio and video into one nested AV latent, which is great for the model and awkward for you. The moment you want to look at, touch, or re-generate one stream - decode just the audio, inspect the video without the audio tangled in, run a second pass on the video only - you're stuck, because most tooling only knows how to handle a single stream. H3 Separate AV Latent is the escape hatch: it takes a native AV latent and splits it into standalone video_latent and audio_latent.

It's the mirror image of the pack's H3 Concat AV Latent, and the two are meant to be used as a pair. Separate to work on the halves, concat to hand H3 back what it wants.

How it works

The split preserves the noise-mask structure of both streams - so if your AV latent came out of a partially-denoised or masked run, the mask stays attached to the right half and you don't silently lose the conditioning state. Beyond that it's clean separation: the nested AV latent is unpacked into the two standalone latents, each of which behaves like a normal ComfyUI LATENT for decoding, saving, or further processing.

The inputs that matter

• av_latent - the native MiniMax H3 AV latent. That's the whole input.

Outputs:

• video_latent - the standalone video stream.
• audio_latent - the standalone audio stream.

How to install

Same pack, same ritual - ComfyUI Turing Utils, "comfyui-svdint4" in Manager:

cd ComfyUI/custom_nodes
git clone https://github.com/wjie98/comfyui-svdint4
cd comfyui-svdint4
python -m pip install -v --no-build-isolation -e ./kernel

Build the kernel once, restart. This node is pure latent plumbing - no models to download.

The honest take

The most common reason you'll reach for this isn't curiosity, it's decode. H3's video VAE decode node expects the video stream, and audio decoders expect audio - so before either of them can run, something has to split the AV latent. This is that something. If your H3 workflow has a "split → decode video" stage, you're already using this node even if you didn't know its name. One tip: if you split, process one stream, and want to put things back together, keep the other half around - don't re-encode from pixels, just concat the latent back. Cheaper and lossless.