Dual VAE Loader (video + audio) ⚡
Two VAEs, one node — the MiniMax H3 video/audio pair
- video_vae
- audio_vae
Some video models need one VAE. MiniMax H3 needs two - a video VAE for the pixels and a separate audio VAE for the soundtrack, and every H3 workflow ends up with a pair of VAELoader nodes sitting side by side, both named vae, both easy to wire to the wrong place. Dual VAE Loader collapses that into a single node with outputs that can't be confused: video_vae and audio_vae. That's the whole job, and it does it well.
What it is
It's a convenience node, not new machinery. Loading is delegated to ComfyUI's stock VAE loader, so nothing behaves differently - taesd entries (the tiny preview VAEs) stay available, offloading still works, and you're not getting some parallel VAE implementation that breaks a week later. What you get instead is a node that says what each output is for, and one less edge case where you decode video with the audio VAE and get garbage and a confused afternoon.
Two inputs, both file pickers:
video_vae_name- the video VAE, e.g.minimax_h3_video_vae_*frommodels/vae.audio_vae_name- the audio VAE + vocoder, e.g.minimax_h3_audio_vae_*.
Outputs are video_vae (VAE) and audio_vae (VAE). Wire the first into your VAE decode and the second into the audio decode. For the other A/V model in this pack (LTX-2.3) the pair is loaded by LTXV23ModelsLoader instead; this node is the H3-flavoured version.
Where it sits in a workflow
It appears in the 🤖 CCTech/GGUF menu category alongside the pack's other loaders. Typical H3 graph: a GGUF-aware model loader (this pack's UNET loader or CCTechClipProjLoader feeding the text encoder) → sampling → CCTechDualVAELoader feeding the video and audio decode nodes. The node is the tidy bit at the end of a graph that is otherwise full of 32B-encoder-sized problems.
Installing it
Same pack as everything else on this page:
cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
cd ComfyUI-GGUF-Loader
pip install --upgrade gguf
Restart ComfyUI (or use Manager and search ComfyUI-GGUF-Loader). No extra dependency beyond gguf for this node - it's just a wrapper around the stock loader. The real requirement is having both H3 VAE files in models/vae/; if either dropdown shows nothing, that's the file missing, not the node.
The one thing worth knowing
Because it delegates to the stock loader, this is one of those nodes where "which VAE is the right one" is still your job - the node just makes it hard to mix them up. H3's audio VAE is not the same shape of thing as its video VAE, and if you swap them the decode will either error or quietly produce nonsense. Load the pair from the same model kit release; mixing a video VAE from one H3 version with an audio VAE from another is a real way to get weird colour or broken audio that looks like a sampling problem. Keep the pair together and this node quietly disappears from your list of things to worry about.
Inputs (2)
| Name | Type | Default | Description |
|---|---|---|---|
| video_vae_name | COMBO | 1 options: pixel_space | |
| audio_vae_name | COMBO | 1 options: pixel_space |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| video_vae | VAE | — |
| audio_vae | VAE | — |