WanVideo Tiny VAE Loader

Video generation is slow enough that watching it happen blind is genuinely painful - a clip can take five to fifteen minutes, and finding out at the very end that the composition was wrong is a bad time. The tiny VAE is the fix. It's a tiny, fast approximate decoder (a TAESD-style model) that turns latents into a rough preview image quickly, so you can watch the generation take shape and kill a bad run early instead of waiting for the full VAE decode. This node loads that tiny decoder from ComfyUI/models/vae_approx.

To be clear about what it is not: this is not your final-quality VAE. It's the preview decoder. The real Wan VAE still does the final decode; the tiny one just gives you cheap glimpses along the way. The KB's LTX section shows exactly how much people rely on this - when LTX 2.3 changed its latent space, "the TinyVAE preview trick stopped working, because the preview decoder was built for the old latent space," and the immediate complaint was "you are generating blind again." That's the whole value: not being blind.

The inputs and output

• model_name (required) - a dropdown of the tiny VAE files found in models/vae_approx. If it's empty, you haven't downloaded one yet - the list is built from files on disk.
• precision (optional, default fp16; also fp32, bf16) - compute precision. fp16 is the sensible default; it's a preview decoder, so precision barely matters for what you're using it for.
• parallel (optional, default false) - a flag for parallel decoding. Leave it off unless you have a specific reason.

The single output is vae (WANVAE) - wire it into the preview/decode point that shows your in-progress frames.

The only input you'll actually touch is model_name. The other two have fine defaults.

How to install it

Comes with Kijai's WanVideoWrapper. ComfyUI Manager: search ComfyUI-WanVideoWrapper, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-WanVideoWrapper
pip install -r ComfyUI-WanVideoWrapper/requirements.txt

then restart. The node needs an actual tiny-VAE file placed in ComfyUI/models/vae_approx - the Wan TAEHV/TAESD-style approximate decoder. Download it, drop it there, restart so the folder is rescanned, and it appears in the dropdown.

Common issues & troubleshooting

Empty dropdown. The most common issue and not a bug: there's no file in models/vae_approx yet. Download the Wan tiny VAE and restart ComfyUI.

Previews look rough or off-color. Expected. The tiny VAE trades fidelity for speed - the preview is an approximation. Judge composition and motion from it, not final color or fine detail; the real VAE decode is the truth.

Previews are wrong for your model. Like the LTX cautionary tale, a tiny VAE is built for a specific latent space. If your previews are garbage while final output is fine, you may have a mismatched tiny VAE. Make sure it's the Wan one, matched to your Wan version.

You wired it as the final decoder. Don't. It'll "work" but your output quality tanks. This is the preview path; keep the full Wan VAE for the final decode.