Anymatix LTXV Audio VAE Loader

If you're generating audio along with video in the LTX family, you need a second VAE - one that lives in audio-land, not image-land - and this is the node that loads it. It looks like a clone of ComfyUI's own LTXV Audio VAE Loader, and that's because it is one: AnymatixLTXVAudioVAELoader takes the exact same checkpoint name and hands it to ComfyUI's LTXVAudioVAELoader internally. The only real difference is the mechanism around it, which is the point of the whole pack.

Why this twin exists

ComfyUI ships native LTX audio nodes in comfy_extras/nodes_lt_audio.py, including an LTXV audio VAE loader that picks from a dropdown of files already sitting in your model folders. The Anymatix version replaces that dropdown with a plain string. Why would you want that? Because the Anymatix app addresses models by URL: a workflow card says "this audio VAE" and a fetcher node downloads it, verifies it, and hands the loader a path that didn't exist when the graph was written. A dropdown can't hold a file that hasn't been downloaded yet; a string can. So this loader takes the fetcher's output and just strips it to the basename before calling ComfyUI's original - the download side and the load side are cleanly separated.

Mechanically, that's the entire trick. Feed it a path, it resolves the filename, ComfyUI's own loader does the rest. Nothing is reimplemented, which is exactly what you want from a loader wrapper - the fewer copies of "load a safetensors into a VAE" there are in the wild, the fewer subtle-drift bugs there are.

The inputs

The single input is ckpt_name, a STRING - the full path to the audio VAE file (or the basename; the loader calls os.path.basename on it anyway, so it lands in the same place either way). In a hand-built workflow you'd type the filename of the audio VAE checkpoint. In an Anymatix-exported workflow, a fetcher node feeds it the downloaded path.

The one output is a VAE, but don't make the mistake of wiring it where your image VAE goes. This is the audio encoder - the thing that pushes audio waveforms into the LTX latent space so the model can generate or condition on sound. It's a VAE the way a bass guitar is a guitar: same noun, different job. Wire it where the audio-capable LTX workflow asks for its audio VAE, alongside the text encoder and audio encoder nodes that sibling pack nodes provide.

Install

This node is part of the anymatix-comfy-nodes pack. ComfyUI Manager: search "anymatix-comfy-nodes" and install, then restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Anymatix/anymatix-comfy-nodes

Restart after. You don't need the Anymatix desktop app to use it, and the pack checks its optional siblings lazily (at call time, not import time) so nothing else breaks if a sibling is missing. The audio VAE file itself needs to live where ComfyUI looks for VAEs - same folder as any other .safetensors VAE - or somewhere your extra_model_paths.yaml points.

Gotchas worth knowing

Because the loader resolves by basename, two audio VAE checkpoints with the same filename in different folders will collide - keep names unique. And the folder matters more than you'd think: ComfyUI's underlying LTXV Audio VAE Loader resolves its filename against the checkpoints folder, not vae - so if you drop the audio VAE .safetensors into models/vae the way you would a normal VAE, you'll get a file-not-found. Put it in models/checkpoints (or wherever your extra_model_paths.yaml points that folder). And remember the shared pack caveat: this pack registers HTTP routes on your ComfyUI server for the Anymatix app's file handling. They're harmless on localhost and the delete endpoints are scoped to ComfyUI's own model roots, but the README asks you to read the list before you run ComfyUI with --listen exposed.