VOID Load VAE

VOID Load VAE is the node you click, set once, and never think about again - and that's exactly its job. It loads the 3D VAE that Netflix's VOID model uses to compress your video into latent space and back, and hands it to the pack's VOID Loader as a custom VOID_VAE type.

If you've built a FLUX workflow you already know the pattern: the loader nodes are the unglamorous plumbing, and the one that bites you is always the one that doesn't use the files you already downloaded.

Why it can't be ComfyUI's built-in "Load VAE"

This is the part that trips people up. VOID is built on CogVideoX-Fun, which uses a 3D temporal VAE (AutoencoderKLCogVideoX) that compresses video across time, not just space - 4x temporal compression along with the spatial downsampling. ComfyUI's stock Load VAE reads the same models/vae/ folder but wraps whatever it finds as an SD-style VAE. That wrapper is useless to this pipeline, so the pack ships its own loader that returns a custom type only the VOID Loader accepts. Don't bother trying to wire a core VAE into it - the sockets don't match, on purpose.

What it does

One input, one output. The vae_name dropdown lists every .safetensors in your ComfyUI/models/vae/ folder, so you just pick the CogVideoX one. Output is a single VOID_VAE that feeds the void_vae socket on VOID Loader.

The one thing worth knowing: the VAE is cached in memory after the first load, so re-running the workflow is instant. It also enables VAE tiling and slicing internally, which keeps peak VRAM down on 12GB cards during encode/decode.

Getting the file

The dropdown reads whatever's in models/vae/, and it'll be empty until you put the file there. Download diffusion_pytorch_model.safetensors (~1.5 GB) from alibaba-pai/CogVideoX-Fun-V1.5-5b-InP (the vae/ subfolder) and drop it into ComfyUI/models/vae/. Renaming it to something like cogvideoxfun_v15_vae.safetensors is fine - the dropdown shows the filename, so pick something you'll recognize.

If you already run CogVideoX workflows you may already have this exact file; if you run FLUX you don't, and no, the SDXL or Flux VAE won't work in its place.

Install

Install the pack once - this node doesn't install separately:

cd ComfyUI/custom_nodes
git clone https://github.com/shanef3d/ComfyUI-VOID

then restart ComfyUI. Or use ComfyUI Manager and search "VOID". The pack's requirements.txt needs diffusers>=0.33.1, which is newer than stock ComfyUI ships, so the first thing it does at load is check your diffusers version and refuse to start if it's too old - more on that in the VOID Loader article.

When it breaks

The failure is almost always a missing file: the node throws FileNotFoundError and prints the exact path it expected and the HuggingFace URL to grab it from. If the dropdown is empty, the file isn't in models/vae/ - it's a folder-name or download problem, not a node problem. If it loaded once and then started erroring after a model swap, restart ComfyUI; the in-memory cache can hold onto a file you replaced on disk.

One honest warning: this node is pure plumbing. If you're here googling it, you've probably already got a workflow half-loaded and just need to know where the VAE goes - and that's the whole answer.