Kandinsky 6 Audio VAE Loader

Kandinsky 6 generates video and audio in one pass. The video half of that is boring plumbing - a Hunyuan Video VAE you load with the normal VAELoader and forget. The audio half is not. There's no stock ComfyUI node for a TOD-VAE plus a BigVGAN vocoder, so this pack ships its own loader, and it has exactly one setting that will ruin your output if you get it wrong.

What it actually loads

Two files, not one. A TOD-VAE checkpoint that turns the model's 40-channel audio latent back into a mel spectrogram, and a BigVGAN v2 vocoder directory that turns that mel into a 44.1 kHz waveform. That's why bigvgan_dir isn't a file picker - it points at a folder that has to contain both config.json and bigvgan_generator.pt. If either is missing, the node throws rather than loading something half-right.

The pack registers ComfyUI/models/audio_vae as its own model folder and also respects extra_model_paths.yaml, so if you keep models on another drive you can point at it instead of copying 1.2 GB around.

The inputs that matter

tod_vae - the .pth/.safetensors/.pt file, picked from that audio_vae folder. The workflow's Download models button drops v1-44.pth there (it's the MMAudio VAE, ~1.2 GB). Don't rename it into something creative; the pack looks in that directory.

bigvgan_dir - the vocoder folder. Default bigvgan_vocoder, containing config.json and bigvgan_generator.pt from NVIDIA's bigvgan_v2_44khz_128band_512x.

mode - currently a single option, 44k. Nothing to choose. It exists because the code path is parameterised by latent downsample factor and sample rate, and 44.1 kHz is what the released checkpoint was trained against.

model_variant (optional) - this is the one you set. base forces scaling factor 0.5302, pretrain / distill forces 0.417, custom uses whatever you typed in scaling_factor. The shipped workflows use base for plain Lite and pretrain / distill for both distilled models.

scaling_factor - the manual value, default 0.5302. The decoder divides the latent by this number before the TOD-VAE sees it, so it's a gain control wearing a math costume. Guess too low and the waveform comes back hot and clipped; too high and your dialogue is a whisper under the ambience.

One VAE output. Wire it into Kandinsky 6 Audio VAE Decode's audio_vae input. That's the whole interface - no video path, no encoding.

Install

ComfyUI Manager → search Kandinsky T8 (publisher t8star). Or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-Kandinsky-T8.git
cd Comfyui-Kandinsky-T8
python -m pip install -r requirements.txt

Use ComfyUI's own Python environment, then restart. You need ComfyUI 0.39.0+, comfy-kitchen 0.2.37+ (it's in requirements.txt), and an NVIDIA card with INT8 tensor cores. If you also have the official Kandinsky 6 extension installed, disable it - both packs register the same node names and you'll get a pile of duplicate-node warnings and unpredictable wiring.

Then get the files. Easiest is the Download models button on one of the pack's workflows: it pulls the shared components from pinned upstream repos and checks size and SHA-256 before use. Manually, the two you need here go to ComfyUI/models/audio_vae/.

Where it goes wrong

The most common one is a scaling mismatch: choosing base while running a distill checkpoint, or leaving custom at 0.5302 on a distilled run. Symptom is audio that's either cracking or inaudible while the video looks fine, which sends people hunting through the sampler instead of this node. Match the variant to the DiT file: Lite → base, Lite distill and Pro distill → pretrain / distill.

Second: pointing bigvgan_dir at the folder parent, or at a vocoder that isn't the 44 kHz 128-band 512x one. The node will tell you the directory is missing its config or generator file - read that message, it's accurate.

Last, a boring one worth knowing up front: the TOD audio decoder is CC BY-NC 4.0 per the pack README. Fine for messing about, not fine for a monetised product. The rest of the pack is MIT.