AuK VAE Loader

Every diffusion model needs a decoder, and usually you don't think about it. AuK gives you a reason to: this VAE is the reason your volume edits survive, and picking anything else in this slot breaks the graph outright.

It's a BigVGAN-style flow codec - 24 kHz, 64 latent channels - and it is the only one of the three model files that ships unquantized, computing in fp32 no matter what precision you chose upstream. That's deliberate. The pack's decode path is custom: AuK Generate / Edit decodes through it directly because stock VAE Decode Audio normalizes loud outputs in the ComfyUI builds the author tested. If you ask for "+5 dB" and the decoder hands you a normalized file, you just spent a generation to change nothing. So the VAE keeps amplitude intact, and you keep the edit.

Inputs and output

One input, one output, no knobs:

• vae_name - dropdown from ComfyUI/models/vae/. There's exactly one file to download here, auk_vae.safetensors, 0.637 GB.
• Output is a core VAE socket, wired into AuK Generate / Edit (.vae).

The type is core VAE, but the sampler checks what's behind it: it wants the object this node produces. Reach into your existing workflow and grab some other VAE - a Qwen image VAE, an SDXL VAE, whatever's in the list - and Generate / Edit stops with "requires the VAE from AuK VAE Loader." That's not being precious; a 64-channel 24 kHz audio latent has nothing in common with an image latent.

Install

ComfyUI Manager → ComfyUI-AuK, or:

cd ComfyUI/custom_nodes && git clone https://github.com/Saganaki22/ComfyUI-AuK
cd ComfyUI-AuK && python -m pip install -r requirements.txt

Then download auk_vae.safetensors from drbaph/AuK-comfyui into ComfyUI/models/vae/. Restart ComfyUI so it rescans the folder - the README calls this out, and it's the usual reason a fresh file doesn't show in the dropdown.

Who touches this VAE, and why that matters

Two places. Generate / Edit encodes your source audio into a reference latent through it (downmixed to mono and resampled to 24 kHz at the integration boundary, then sampled with the run's seed), and it decodes the final latent back to a waveform, trimmed to the exact number of samples seconds asked for. So the VAE sits on both ends of every edit, and both ends affect what you hear.

The seed point is worth knowing because it explains non-determinism people attribute to the diffusion side: the seed drives the reference-latent sampling and the noise. Change it and you change both.

Troubleshooting

"Requires the VAE from AuK VAE Loader." Covered above - you wired the wrong VAE, or you loaded the model with a core VAE loader.

The dropdown shows the file as missing after loading someone else's workflow. On Windows, ComfyUI lists subfolder models with backslashes, so a path saved on a Linux machine won't resolve. Re-select auk_vae.safetensors and re-save the workflow.

Tiled VAE nodes do nothing for you. The adapter implements standard untiled audio encode and decode only. There's no spatial tiling story here, so don't plan on VAE tiling to rescue a long clip - keep the clip shorter instead.

Output sounds loud or clipped anyway. That's the generation, not the decode - the codec is preserving what it was given. Run seconds=0 for gain edits, decode straight to Preview Audio or Save Audio, and skip post-hoc normalization, which will happily undo the whole point.