WanVideo VAE Loader(SE)

This one does exactly one job: load the VAE that turns Wan's latents back into actual video frames. It's the "SE" fork's copy of Kijai's VAE loader, and it's the second node you wire up in this pack after the model loader - everything downstream that decodes or encodes pixels (WanVideoSEDecode, WanVideoSEImageClipEncode) takes its output as a required input.

One thing worth flagging so it doesn't throw you off: the node's own description says it "loads Hunyuan VAE model from 'ComfyUI/models/vae.'" That's not a mistake in your setup - it's a leftover from the code this was forked from. Kijai's ComfyUI-WanVideoWrapper handles more than one video model family, and this loader's description text simply didn't get updated for the Wan-only fork. In practice you're loading a Wan VAE file here, not anything to do with Hunyuan.

How it works

No real mechanism beyond loading a file - you pick a VAE checkpoint and a precision, and out comes a VAE object other nodes decode and encode through. It matters more than it looks: VAE precision affects how much VRAM the decode step needs later, which is the part of a video pipeline that tends to run you out of memory first.

The inputs and outputs that matter

Just one required field, model_name - an enum populated from whatever files sit in ComfyUI/models/vae, and its own tooltip is blunt about it: "These models are loaded from 'ComfyUI/models/vae.'" One optional field, precision (fp16 / fp32 / bf16, default bf16) - bf16 is the sane default and matches what most Wan setups use elsewhere in the pipeline; fp32 is more accurate but heavier, and rarely worth it unless you're chasing a specific artifact.

Output is vae, typed WANVAE, feeding both WanVideoSEDecode (turns your sampled latent into frames) and WanVideoSEImageClipEncode (encodes your start/end images going in).

How to install it

Same as the rest of the pack: ComfyUI Manager, search "ComfyUI-WanVideoStartEndFrames," install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/raindrop313/ComfyUI-WanVideoStartEndFrames

The pack's README doesn't repeat model download instructions for the VAE either - it points you at Kijai's ComfyUI-WanVideoWrapper, since this whole fork is built on his code and shares his model layout. Grab the Wan2.1 VAE file from there and drop it in ComfyUI/models/vae.

Common issues & troubleshooting

Empty model_name dropdown. Same fix as everywhere else in this pack: the VAE file isn't in ComfyUI/models/vae yet, or you added it after ComfyUI already started - restart to pick it up.

Decoded video is noise or garbage. This node won't warn you if you point it at the wrong file - a leftover SD or SDXL VAE sitting in the same folder from an image workflow will load without error and then produce nonsense. Make sure model_name is actually pointed at the Wan VAE, not whatever else happens to be in that directory.

Beyond that, this is about as low-drama as nodes get in this pack - no quantization, no attention backend, nothing to misconfigure once the right file is in the right folder.