Create H3 Audio RefMod

MiniMax H3 generates video with native stereo audio, and this node is how you hand it a sound to reference. Extract H3 Audio RefMod takes an AUDIO input, runs it through H3's audio VAE, and saves a .safetensors audio RefMod you can load alongside visual refs and inject with Apply H3 RefMod - no separate audio pack needed, since v0.2.0 folded audio support straight into this one.

One thing to know before you get excited, and the README says it louder than anything else: speaker-identity transfer doesn't reliably work yet. The author's own tests could not reproduce a reference speaker's voice - a masculine reference on a female character still came out feminine. Music-reference conditioning did produce a working result (background music carried into the generated video), but treat this release as "music, singing style, ambience, SFX" territory and not voice cloning. Keep different voices in separate files anyway, so you can fade them independently when it does land.

How it works

The implementation follows the tested ComfyUI-H3AudioMod encode path: audio is normalized to stereo, encoded to [1,32,2,T] latents at 40 frames/second and 2 tokens per frame, in bounded 10-second encode chunks. So token cost scales with duration - 2 × T. max_seconds (default 30) explicitly selects the clip's initial duration; chunk boundaries can nudge codec continuity, so this is reference conditioning, not a promise of lossless audio.

Inputs are a short list and mostly self-explanatory:

• audio - the AUDIO clip (from any audio loader/preview node that outputs AUDIO).
• audio_vae - the H3 audio VAE (the 32 kHz codec), not the video VAE. Connecting the wrong one is the #1 error here.
• name / subfolder / description / concept_type - naming and metadata. concept_type is one of voice, singing, music_style, sound_fx, ambience - descriptive, stored in the file, and it feeds the loaders' prompt_hint.
• max_tokens (default 5120) with budget_policy - on overflow, error raises; truncate keeps a contiguous prefix rather than resampling audio in latent time.
• save - off if you're piping straight into Save H3 RefMods and don't want a double write.

The single output is an H3_REF_MODS bundle with the audio mod at strength 1.0 - feed it to Apply H3 RefMod like any visual mod, or load the saved file later with Load H3 RefMods.

Install

ComfyUI Manager (search "ComfyUI-MiniMaxH3Mod") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Luisacaotica/ComfyUI-MiniMaxH3Mod

Restart; light deps only. You need the H3 weights and its audio VAE already working in a current ComfyUI - remember the H3 Community License excludes the US, EU, UK and Korea from the local-weights path.

Common issues

Beyond the voice-transfer caveat, the usual stumbles are wiring the video VAE into audio_vae, clipping a track longer than you meant (watch max_seconds - it sets the duration, and 30 seconds of audio at 2 tokens/frame is real budget), and expecting an audio mod to also control the visuals. It doesn't - audio and visual refs are separate files with separate budgets, which is why the Master node exists if you want both from one run. No <Name> prompt triggers are implemented here; the prompt is your only text-side handle, and the README is honest that it can't force selective transfer.