VAE Loader Video+Audio

Video models do something weird: some of them come with two decoders. LTX2 (and its LTX-Video ancestors) is a text-to-video model that also generates sound, which means it needs a VAE for the video frames and a separate audio VAE - essentially a vocoder - for the audio track. This node loads both from your models/vae/ folder in one go and puts them on two sockets, video_vae and audio_vae.

Why a dedicated node instead of two stock loaders? The short answer is the GGUF LTX2 flow. GGUF-quantized diffusion models are pure inference weights - no VAE baked in, no CLIP, no nothing but the transformer. So when you run LTX2 as GGUF (which is the whole reason people run GGUF: half the VRAM of fp16), you're forced to supply both decoders yourself. This node exists precisely so that setup is two dropdowns instead of a scavenger hunt.

The pack's own words: "useful for GGUF LTX2 where neither VAE is baked into the model file." That's the use case, and it's honestly about the only one. If you're on a full LTX2 checkpoint with everything baked, you don't need this node at all.

Inputs

• video_vae - the video/image VAE file. Set to None to skip and wire a video VAE from elsewhere.
• audio_vae - the LTXV/LTX2 audio VAE (look for the vocoder-type weights in your vae folder). None skips it too.
• disable_offload - default true, keeping both VAEs resident on the GPU. Video decode is expensive; leave it on unless VRAM forces your hand.

Each socket is independent, so you can load just the video VAE here and pull audio from somewhere else, or vice versa. Both files resolve through the same folder-role validation as every loader in this pack, which means path and format checks happen before anything is deserialized.

Outputs

• video_vae → the video VAE decode path
• audio_vae → the audio decode path

Straight wires, nothing else - which is exactly the point of a focused dual loader.

Installing

Pack-standard:

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

or ComfyUI Manager → ComfyUI Smart Model Loader, restart.

Common issues

• The audio VAE won't load / shape errors - you've grabbed a video VAE in the audio slot. LTX audio VAEs are vocoder weights, not the same file family, and the node (like ComfyUI generally) won't silently make a mismatch work.
• "None" everywhere and no decode - with both slots skipped this node does nothing by design; it's a convenience node, not a router.
• Video VAE file isn't in the list - it must live under models/vae/; subfolders work, but it won't scan arbitrary locations.