VAESave

You'll use this node about twice a year, and both times you'll be glad it's a one-liner. VAESave takes any VAE you have loaded and writes it to disk as a standalone .safetensors in your ComfyUI output folder. That's the whole job - but it fills a real gap, because VAEs usually arrive stuck inside a checkpoint, and sometimes you want one on its own: you downloaded a checkpoint just for its nicer VAE, you merged or tweaked a VAE, or you want to stop loading a multi-gigabyte checkpoint every time just to get its small VAE component.

How it works

Under the hood it's the sibling of ModelSave and CLIPSave: the node grabs the VAE's state dict (vae.get_sd()), writes it as a safetensors file, and embeds the generating prompt and workflow as metadata so the file keeps its provenance. It's an output node, so it has no outputs and doesn't need to be wired into anything downstream - you run the graph and it saves.

The inputs

• vae - the VAE to save, usually from a VAELoader or the VAE output of a CheckpointLoaderSimple.
• filename_prefix - a path + name relative to the output folder, default vae/ComfyUI_vae.

Run it and you get output/vae/ComfyUI_vae_00001_.safetensors, with the counter auto-incrementing so you never clobber an earlier save. VAEs are small - a few hundred megabytes versus the gigabytes of a full model - so this is effectively instant.

Common issues

• Can't see the saved VAE in the loader. It's in output/vae/, and loaders scan models/vae/. Copy it over and refresh the dropdown - or just keep it where it is and reference the path directly.
• Garbled or wrong-colored output after using the saved VAE. The VAE you extracted doesn't match the model's latent space. A VAE is architecture-specific: SD1/SDXL/SD3 and the video models all use different latent spaces, and swapping them gives you noise or tinted garbage. Match the VAE to the model family.
• "It saved nothing" confusion. Output nodes have no outputs - that's expected. Check the output folder for the numbered file.

VAESave has been in ComfyUI core since October 2023, one of the original save nodes. Nothing to install; it's the boring, reliable piece of the merging family that lets the stuff you export actually live on as files instead of evaporating when you clear the graph.