MiniMax H3 Audio Ref Extract

The thing that makes MiniMax H3 different from the silent-video models it competes with is that audio is generated jointly with the picture rather than bolted on afterwards - dialogue, room tone and all. The flip side of a unified context is that you can give it audio as context: feed it a clip and it can carry that character's voice into the generated output.

UC_MiniMaxH3AudioRefExtract is the small, single-purpose half of that workflow. It takes one audio clip, encodes it through the H3 audio VAE into a reference latent, and hands you a MINIMAX_H3_REF you can save, reuse, and apply to conditioning.

What it does, concretely

audio takes one mono or stereo clip as ComfyUI's AUDIO type. Mono gets upmixed to stereo internally, because H3's audio VAE expects a stereo pair - you don't need to do anything about that.

audio_vae is the critical input and the one people get wrong: connect the MiniMax H3 audio VAE, not the video VAE. They are different encoders for different modalities; the wrong one fails instead of quietly producing something plausible. If you're not sure which is which in your workflow, check the loader it's coming from - the audio VAE is the one that isn't feeding your visual latents.

The result is a 32-channel audio latent stored inside the ref object, and description (multiline, default empty) is optional text saved alongside it. The description does not change your prompt and does not act as a trigger word. It's a label for your own future self.

Outputs are ref and info. The info string summarises the reference, including its latent shape and token cost - for audio that's twice the latent length, which is a much cheaper reference than a video one and often worth comparing numbers before you decide what to carry.

What it is not

This is not voice cloning, and the node says so in its own tooltip: "Audio references do not guarantee the same voice in generated output." What H3 does with a reference is conditioning, not identity transfer with a guarantee attached. If you need reproducible voice identity, that's a different toolchain - and H3's own licensing situation (the Community License excludes the US, EU, UK and South Korea from its Applicable Territory) is worth being clear about before you point this at anything commercial.

It's also not a learning node. Nothing here trains, fine-tunes or adapts anything. It encodes and packages.

Install

Manager → search "ComfyUI-UtilsCollection", or:

cd ComfyUI/custom_nodes
git clone https://github.com/silveroxides/ComfyUI-UtilsCollection
# restart ComfyUI

No models download for this node - it uses the H3 audio VAE your workflow already loads. The pack's requirements.txt brings in opencv-python, huggingface_hub, unifiedefficientloader and typing-extensions, and the collection imports kornia when it loads without declaring it. So if the node list shows nothing from this pack after a fresh install, install kornia into ComfyUI's environment before you start debugging anything else.

You'll want the rest of the family with it: MiniMax H3 Ref Save to write the ref to disk, MiniMax H3 Ref Load to bring it back, and MiniMax H3 Ref Apply to append it to your H3 conditioning. Extraction is cheap enough to do inline, but the reason this suite exists is that you shouldn't have to - a saved reference is one load away instead of one VAE pass away.

Getting it right

The practical trick is to attach this to the same conditioning your prompt already uses and let the reference do its work, rather than trying to describe the voice in text. And keep the audio reference short and clean: a couple of seconds of the voice you want is a reference. A three-minute scene with music underneath is a reference that fights your prompt.

If the encoded reference surprises you - the info output is your instrument here, since it prints the latent dimensions and the token cost before anything generative happens. A reference that's much bigger than you expected is one that's about to eat your context window, and the audio VAE won't trim it for you.