VAE Loader Video+Audio

Video generation with native audio means two decoders, not one: a video (or image) VAE that turns latents into frames, and an audio VAE that turns them into sound. Two loaders for that is two dropdowns, two nodes and a graph that's 30 pixels taller than it needs to be. VAE Loader Video+Audio [SML2] puts both on one node with separate outputs.

Why this exists now

The 2026 video models went omni-modal - LTX-2 and MiniMax H3 both generate sound alongside frames rather than bolting an audio pass onto a silent generator. That means two VAE files in play, and audio VAEs are awkward customers: they frequently ship inside a checkpoint as audio_vae. and vocoder. prefixed weights rather than as a tidy standalone file. ComfyUI's own audio VAE loader filters and remaps those prefixes to make a VAE out of them. This node is the two-slot version of that idea, for when your graph wants both on the same wire junction.

How it works

Three inputs, all with dropdowns over your models/vae folder:

• video_vae - the video or image VAE. Default None, which skips it.
• audio_vae - a ComfyUI-supported audio VAE, such as MiniMax H3 or LTX. Default None, which skips it.
• disable_offload - defaults to true, requesting a full load for each selected VAE.

None is a real choice, not a placeholder. Skip a slot and its output is None rather than an error, so you can use this node as a single VAE loader with a spare socket, or ship a workflow where the audio side is optional.

Both slots go through the same external VAE load path the single loader uses - metadata preserved, Qwen Image 2.1 Diffusers-format handling included, tensors validated. The consequence is worth stating plainly: an audio VAE loads here only if your ComfyUI build recognizes its keys. That's what "ComfyUI-supported" in the tooltip means, and it's the difference between this node being convenient and this node being useless on an older install.

Outputs are video_vae and audio_vae, both VAE type. Wire the first to your decode/tiled-decode node and the second to whatever handles the audio track in your pipeline.

When to use this instead

The pack has a boundary here and it's worth respecting. These standalone VAE nodes are plain file dropdowns. If you want per-file Source / integrity metadata, an expected SHA-256, or a Download Manager return target that fills the field after a verified download, you use Smart Model Loader's VAE or Audio VAE sections instead. Convenience and provenance are separate features in this pack; pick the one you need.

Similarly, if your workflow is image-only, the single VAE Loader is the smaller node. This one exists for the GGUF/LTX-style flows where the video VAE and the audio VAE are genuinely separate files and you're tired of hunting for both.

Installing it

One install for the whole pack. ComfyUI Manager → search Smart Model Loader Nova, or:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_SmartModelLoader_Nova.git
cd ComfyUI_SmartModelLoader_Nova
python -m pip install -r requirements.txt

Run that with ComfyUI's Python environment, then restart. No Nunchaku, no PuLID, no GGUF backend needed for this node - just the base requirements. Files go in ComfyUI/models/vae.

Where people get burned

Audio VAE files are not interchangeable and often mislabelled. A file that looks like a checkpoint may be an audio VAE with the right prefixes, or may be neither. If the dropdown selection loads but downstream audio nodes reject it, suspect the file before the node.

None doesn't warn you. Skipping the audio VAE and then wondering why there's no audio track is the most likely beginner mistake here. The output is quietly None; nothing pops up to tell you.

Offload is a request, not a promise. As with the single VAE Loader: disable_offload asks for a full load, but ComfyUI still decides device placement and residency. It won't pin a 1GB audio VAE in VRAM forever just because the checkbox is ticked.

Licensing is on you, not the loader. If you're running MiniMax H3's open weights, note that its community licence excludes the US, EU, UK and Korea from the applicable territory. That's a legal constraint on running the weights, and no node in ComfyUI will enforce it for you.