H3 Continuum Reference Audios

MiniMax H3 is omni-modal in the way that actually matters for video: it treats audio as part of the same generation context, so you can hand it a voice clip and ask for a character who sounds like it. The H3 Continuum Sampler V3.8 gives you three separate Reference Image slots but only one audio bundle socket - and this node is the box that turns up to three standalone audio files into that single Audio References (Optional) connection.

Call it what it is: a bundle helper, not a sampler and not a loader. You connect audio files, it packs them, and the sampler handles the rest.

The inputs

Four optional sockets, and the one rule is order:

• reference_audio_1, reference_audio_2, reference_audio_3 - connect in order, without gaps. Don't wire slot 1 and slot 3 while leaving slot 2 empty; the ordering is the whole point.
• reference_audio_vae - one shared Core Audio VAE that encodes every connected reference. The tooltip says it plainly: required when any audio is connected.

The ordering isn't decoration. Connected inputs map directly to <Audio 1>, <Audio 2> and <Audio 3> in your prompt, so your prompt and your wiring have to agree. Audio 1 in the prompt is the file in slot 1 on the node.

The output

A single audio_references value of the custom H3_CONTINUUM_AUDIO_REFERENCES type, which plugs into the sampler's optional audio_references socket. Don't try to reuse that custom-typed output anywhere else - it's encoded for the sampler, full stop.

The contract that trips people up

This is reference audio, meaning conditioning only. These clips influence voice, timbre, sound style - they never become the finished video's audio track. The pack is emphatic on this: "generated final audio is never replaced." If you want a source track preserved as the output sound, that's the separate Driving Audio path on the sampler, not this bundle. And don't connect this bundle together with the sampler's legacy single Reference Audio (Optional) socket - pick one path.

Installing

It's the one-pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git

Restart ComfyUI, or grab it via ComfyUI Manager under "H3 Continuum". No extra pip dependencies, and the loaders feeding this node's slots come from the same pack.

One honest note: multi-reference audio conditioning is new territory, and the pack's own docs describe subjective listening tests for boundary artifacts rather than guarantees. Start with a single reference to hear how H3 responds before you stack three voices and hope the prompt sorts them out.