Vae Selector

The Vae Selector is the least glamorous node in the ComfyUI-Info-Prompt-Toolkit pack, and that's fine - it's a utility, not a headliner. It exists for one reason: if your generated images come out grey, washed out, or just flat, you're almost certainly missing the right VAE, and this node makes picking one a dropdown instead of a trip back into the file browser. It's the ComfyUI equivalent of the "just set vae-ft-mse-840000-ema-pruned" advice that has been handed out on every Stable Diffusion forum since 2023.

What it actually does is simple: it lists every .safetensors file sitting in ComfyUI/models/vae and hands the one you pick downstream. No downloads, no API calls, no models fetched at runtime - the pack's own README is careful to point out that selectors like this are pure routing. Everything the node needs is already on your disk.

Why it lives in a "toolkit" pack

The useful part isn't the dropdown itself. It's that this pack wraps generation settings into a single image_info bundle that gets serialized into A1111/Civitai-compatible metadata when you save. A VAE choice baked into that bundle means the next run - yours, or whoever you share the image with - can reproduce the exact color pipeline instead of silently using whatever baked VAE the checkpoint shipped with. In a pack philosophy where the whole point is "carry your trial results into the next pass," a selector that remembers its choice is the connective tissue.

Inputs and output

The node has exactly two inputs and one output:

• vae (required) - the dropdown of VAEs found in your models folder. Tooltip: "Select VAE."
• sha256 (optional) - a cached hash used by the right-click View Model Info... action as a fallback for matching the VAE on Civitai. You can usually leave it alone.
• vae (output) - the selection, which you wire into the pack's Image Info Context, Load New Model, or Use Loaded Model inputs, or anywhere else that wants a VAE reference.

That's the whole surface. There's no "wrong" setting on this node; the only meaningful choice is which VAE you pick.

Installing it

The node ships in ComfyUI-Info-Prompt-Toolkit. Easiest path is ComfyUI Manager - search the pack title and hit install. Manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/kinorax/comfyui-info-prompt-toolkit.git
cd comfyui-info-prompt-toolkit
pip install -r requirements.txt

Then restart ComfyUI. Note the pack is built on ComfyUI's newer node API, so it wants a fairly recent ComfyUI (≥ 0.17.0) - if the node doesn't show up after install, update ComfyUI first.

Where people get burned

The classic failure isn't this node at all - it's expecting it to fix a missing VAE. If the dropdown is empty, you have no VAEs in ComfyUI/models/vae, so go grab one (for SD 1.5, vae-ft-mse-840000-ema-pruned is the standard). And a caveat worth repeating from the ecosystem: a VAE from the wrong architecture - say, an SDXL VAE on a Flux workflow - won't look subtly off, it'll produce noise or flat color. The channel counts don't match. So when your images look wrong after picking a VAE, question the pick, not the node.