TurboWan VAE Loader

Wan is a video model, so it doesn't use a normal image VAE, and that's the whole reason this node exists. ComfyUI's standard VAELoader loads 2D VAEs that operate on (B, C, H, W) images. Wan's VAE is a 3D causal video VAE that compresses the temporal dimension too, working on (B, C, T, H, W) - it's what turns your 77 frames into ~20 latent frames and back. Feed the wrong VAE to the sampler and the shapes don't line up and nothing works.

TurboWanVAELoader loads the Wan 2.1 video VAE through the pack's vendored Wan2pt1VAEInterface, so the VAE it outputs speaks the language the TurboDiffusionI2VSampler expects.

What you need to download

One file, from the same HuggingFace repo as the diffusion models:

• wan_2.1_vae.safetensors (or the .pth equivalent, ~2.5 GB)

Put it in ComfyUI/models/vae/. The node's vae_name dropdown scans your vae folder (plus diffusion_models/ and the pack's own checkpoints dir, with a lazy fallback to wan_2.1_vae.safetensors if it finds nothing). It's a lazy loader like the model loader - nothing actually loads until the sampler needs to encode your start image.

Using it

One required input (vae_name), one output (vae). Wire the output into the sampler node's vae socket and forget about it. You don't pick a checkpoint type, you don't mess with dtype - the node handles the 16-channel latent layout and the temporal compression for you.

A couple of honest caveats:

• Don't grab "any VAE" - it has to be the Wan 2.1 video VAE. A standard SD/SDXL VAE will load and then fail at encode time with a shape error, because it can't see the time dimension.
• The dropdown looks for *vae* in its filenames. If your file is named something unusual, the node won't list it. Rename it to contain "vae" and it shows up.
• VRAM is a non-issue here - the VAE moves to the GPU to encode/decode and returns to CPU afterward, and the sampler aggressively empties the cache around it.

If you're using the pack's other route (the TurboWanSampler + KSamplerAdvanced path), you can feed this VAE there too - it's the same interface ComfyUI's own Wan nodes use.