VELVET VICE MiniMax H3 — AUTO VAE Loader

Here's the thing most people discover the hard way about MiniMax H3 in ComfyUI: it doesn't have one VAE, it has two. The omni-modal trick that makes H3 special - video and native stereo audio generated as one context - means the video latent and the audio track come out through separate autoencoders. The VELVET VICE MiniMax H3 AUTO VAE Loader is the node that loads whichever one you need, and it picks the right file for you.

The split is exactly what you'd expect from the filenames:

• video (default): minimax_h3_video_vae_fp16.safetensors - decodes the actual frames.
• audio: minimax_h3_audio_vae_fp32.safetensors - decodes the synced audio track out of its latent.

You set vae_role to tell it which side you're after. In the full Velvet VICE H3 workflow this distinction matters constantly: the pack's MUTED · VIDEO ONLY mode deliberately never loads the audio VAE at all, because why pay VRAM for a decoder you're not going to use.

The mechanism is the same pattern as the pack's other loaders: vae_name defaults to AUTO, which resolves the official basename even when the file is parked in a subfolder of ComfyUI/models/vae/, then hands off to ComfyUI core's VAELoader. Nothing magical - but the AUTO basename resolution is genuinely useful, because the "official MiniMax H3" folder names from various download sources disagree with each other, and this shrugs past that.

The one output is a standard VAE socket. Wire the video VAE into your VAEEncode / VAEDecode and the H3 image-to-video nodes; wire the audio VAE into the audio decode path only if you're rendering WITH SOUND.

A rule that catches people

ComfyUI-GGUF has no VAE loader, so there is no such thing as a GGUF H3 VAE. The VAE layer of this pack is native safetensors only - the System Hub will refuse a GGUF video VAE and fall back to "MUTED / VIDEO ONLY" territory for audio. Don't hunt for quantized VAE files; there isn't supposed to be one.

Install and files

The whole pack installs together - no per-node install:

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-minimax-h3

Restart ComfyUI afterwards (or use Manager and search "velvet-vice-minimax-h3"). There are no pip dependencies to fight with; this is pure wrapper code over core.

Both VAE files come from the MiniMax H3 weight release and live in ComfyUI/models/vae/:

minimax_h3_video_vae_fp16.safetensors
minimax_h3_audio_vae_fp32.safetensors

AUTO finds them in subfolders too.

When it goes wrong

The classic error is a role mismatch: you ask for audio but only dropped the video VAE in, and the resolver finds nothing. Check the file is actually in your vae folder before blaming the node. And if you're running a MUTED workflow, don't fight it - the pack is telling you the audio VAE genuinely isn't needed, and leaving it out is a legit way to claw back RAM on a 40GB-class model. That's not a bug, it's the design.