AuK VAE Loader X

Why you'd reach for it

Audio diffusion models work in a compressed latent space, same as image models: the model denoises a latent, and a decoder turns that latent back into a waveform. AuK's decoder is auk_vae, and this loader is the one that mounts it.

There are two reasons it gets its own node rather than being folded into the model loader. The first is that the VAE is shared - one VAE can serve as many generations as you like, and swapping Base for Flash doesn't mean re-loading the decoder. The second is more interesting and it's stated right in the tooltip: decoding through this VAE preserves AuK's volume edits. Volume changes, speed, pitch - those are baked into the latent as the model sees it, and a different decoder wouldn't reproduce them faithfully. This is why you shouldn't be tempted to route AuK latents through some other audio VAE.

The rule for this node is one line: always the original unquantized AuK VAE, and it always computes in fp32. No bf16, no int8. The 0.64 GB file is small enough that there's no reason to quantize it, and every reason not to - the VAE is the final arbiter of how everything sounds, including the edits you asked for.

How it works

vae_name is a combo drawn from ComfyUI's VAE folder, so drop auk_vae.safetensors in ComfyUI/models/vae/ and it shows up after a restart. One output: VAE, which goes to AuK Generate / Edit. That's the only consumer in the suite.

If you're coming from the image side, note the difference in shape: there's no VAE Encode/Decode pairing here, and no tiling. AuK's encode happens inside the instruction/generation pipeline; this node just supplies the decoder's weights.

Install

ComfyUI Manager → search WorkflowX Configurator, or:

cd ComfyUI/custom_nodes
git clone https://github.com/haroonaslam/WorkflowX-Configurator
cd WorkflowX-Configurator
pip install "transformers>=5.3,<6" safetensors PyYAML tqdm

That dependency line is the AuK optional set. The base pack (PyWORLD, numpy, scipy) is for the Voice ChangerX node and isn't needed for these, but installing it costs nothing.

Then:

ComfyUI/models/vae/auk_vae.safetensors   (0.637 GB)

from the drbaph/AuK-comfyui repo. Restart ComfyUI after adding it - model folders are scanned at startup, so a file dropped in while the server is running won't appear in the dropdown.

Where it goes wrong

"Missing VAE" in a loaded workflow. Usually a filename or folder mismatch, occasionally the Windows subfolder path quirk the AuK docs call out - re-select the value from the dropdown and it sticks.

A quantized or foreign VAE. There's no guard here; the loader will accept anything in the folder and the failure shows up as audio that sounds wrong rather than as an error. If your volume or speed edits come out muted, subtly off, or just don't land, check that you're on the unquantized AuK file before you start re-reading your instructions.

Don't conflate this with the audio VAE in the H3 RefMod suite. They're different models for different jobs that happen to share a folder and a type name. AuK's VAE decodes speech latents; H3's decodes reference clips for the video model. Mixing them produces errors that read as if the model is broken.