MiniMax H3 AUDIO VAE (CRT AutoDL)

MiniMax H3 generates sound with the picture - dialogue, SFX, room tone, jointly, not as an afterthought pass. That requires a VAE that can turn audio latents back into waveforms, and MiniMax H3 AUDIO VAE (CRT AutoDL) is the node that provides it: no inputs, one output, and the weights download themselves on first run.

It's part of PGCRT's CRT-Nodes AutoDL family, the pattern where every model is pinned to a known URL and fetched into the right folder automatically. This one pulls minimax_h3_audio_vae_fp32.safetensors from Comfy-Org's MiniMax-H3 repo into ComfyUI/models/vae/. Note the fp32 - the audio VAE is kept at full precision even though the video VAE ships fp16. That's deliberate on Comfy-Org's part; audio decode is cheap relative to video, and full precision avoids quantization artifacts in the waveform.

Where it fits

H3 splits its VAE into two: a VIDEO VAE for the frames and this AUDIO VAE for the sound, and a complete H3 graph uses both. The audio VAE decodes the audio-track latent - the same one the model generates alongside the video - into an AUDIO tensor you can preview, save, or run through a Save Audio With Path (CRT) node if you want the soundtrack as its own file.

Output is a single VAE, wired into whatever decodes the audio portion of your H3 latent. If your workflow is silent video only, you can technically leave this node out - but then you're paying for the omni-modal model and using half of it, which is its own waste.

Installing it

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

or ComfyUI Manager → search CRT-Nodes, restart. First run downloads the VAE; at fp32 it's a few GB, so the download is the wait, not the node itself.

Common issues

• No audio out - check that the audio latent is actually being decoded with this VAE, and that you've also got the VIDEO VAE on the video side. Mixing H3's audio VAE with another model's latents is nonsense; they're paired.
• License - H3 weights are under the MiniMax H3 Community License, which excludes the US, EU, UK, and Korea from the applicable territory. The VAE ships under that umbrella too.
• Update breakage - red socket or NaN after a CRT update → right-click → Fix node (recreate).