VAE Selector

Unlike its sibling Checkpoint Selector, this one does the real work. HelperNodes_VAESelector actually loads a VAE and hands you a usable VAE object - the thing you wire straight into VAEDecode to turn latents into pixels.

It's a plain dropdown of everything in your ComfyUI/models/vae folder. Pick vae_name, and the node loads the file and outputs the decoded VAE. The encoding happens at graph-execution time, so it costs a bit of load time per run, but it means the choice rides a wire and Save Image can log which VAE you used... well, actually, Save Image doesn't record the VAE. But VAEDecode gets its input from a clean, explicit wire instead of you hoping the right file is in the right default folder.

How it works

The node builds its dropdown by listing folder_paths VAE files plus the vae_approx folder. The clever part is the approximate-VAE handling: if you have both the encoder and decoder for taesd or taesdxl installed, they show up in the list as pseudo-VAEs, and the node assembles them from their parts automatically. Pick one of those and you get a tiny, blazing-fast approximate VAE - genuinely useful for quick previews before committing to the full-quality decode.

Inputs and outputs

• vae_name - one required enum: your VAE files, plus taesd/taesdxl when present.
• VAE output - a real loaded VAE. Wire into VAEDecode's vae input.

The honest take

Two opinions, both from hard-won experience. First: the right VAE is the one your checkpoint was trained against, not the one with the best reputation. Modern checkpoints mostly bake theirs in, and on those you don't want to bolt on an extra VAE at all - attaching one is now more often the mistake than the fix. The standalone-VAE ritual is a legacy of SD 1.5, where the 840k file fixed washed-out colours. Second: taesd and taesdxl are great for previews and genuinely bad for final output - use them to check a batch cheaply, then swap to the real VAE for the keepers.

Install

Search ComfyUI-Helper-Nodes in ComfyUI Manager and install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/teward/ComfyUI-Helper-Nodes

Restart afterwards. Requirements are numpy, pillow, piexif, whratio - nothing exotic, no model downloads. The VAE files themselves are your existing ones, wherever you already keep them.

Troubleshooting

If taesd/taesdxl don't appear in the list, you're missing one of the encoder or decoder files in models/vae_approx - both halves are required. If the dropdown is empty, your models/vae folder is empty. And remember the approx VAEs are for previewing; if your final images look soft, you're probably decoding through one of those.