PixlStash VAE Loader

The VAE is the quiet workhorse: it turns latents into pixels, and on a modern model the difference between a good VAE and a bad one is the difference between crisp detail and mushy garbage. If you run PixlStash - the self-hosted image manager from Pikselkroken - this node is the drop-in replacement for ComfyUI's built-in Load VAE that picks the file off the PixlStash model shelf instead of a dropdown.

What it is

One input, one output: vae_sha256 in, vae out, wire it into a VAE Decode (or VAE Encode) exactly as you would from the built-in. The Browse VAEs… button opens a grid of the VAEs on your shelf and writes the SHA-256 into the widget - you never type there. It's shaped after the built-in VAELoader on purpose, so it drops into a graph where a built-in already sits.

How it works

The node resolves the file from the shelf: used in place when ComfyUI shares PixlStash's filesystem, or fetched once, verified against its SHA-256 before anything is written, and cached at <your first vae directory>/pixlstash/<sha256>.safetensors for every later queue. It also does something the built-in won't: it asks the loaded VAE to validate itself (throw_exception_if_invalid), so "this file isn't actually a VAE" becomes an error right here instead of a tensor-shape crash three nodes downstream. Small touch, genuinely useful when you've grabbed the wrong file off the shelf.

One deliberate gap: TAESD and the other vae_approx previews are not on the shelf. Those are pairs of encoder/decoder files that live in ComfyUI's own vae_approx folder, and PixlStash's scanner excludes them from the vae kind on purpose - they're not real autoencoders. So don't go hunting for your preview VAE in the Browse grid; for those, use ComfyUI's own Load VAE as you always have. On a typical text-to-image run you want the real VAE here anyway.

Install and setup

cd ComfyUI/custom_nodes
git clone https://github.com/Pikselkroken/ComfyUI-PixlStash.git

Or search ComfyUI-PixlStash in ComfyUI Manager. Dependencies are just requests and Pillow, both already in any ComfyUI - no heavy install, no model downloads. Restart, then put your PixlStash URL and API token under Settings > PixlStash. The shelf routes are OWNER_ONLY, so you need an owner token here - the resource-scoped share token that works for the Picture Loader returns 403. Multi-user ComfyUI (--multi-user) isn't supported.

Where people get burned

• Expecting a preview VAE in the grid. Not there by design; use the built-in Load VAE for TAESD and friends.
• A VAE alone doesn't run a model. For a bare diffusion model you'll be pairing this with the PixlStash Checkpoint Loader and CLIP Loader anyway - the VAE output wires into decode, the CLIP handles the prompt.
• Shelf fetch needs the v1.10+ server. On an older PixlStash the node only works when ComfyUI and PixlStash share a filesystem, since the shelf routes that serve file bytes came with the model-shelf release.

And the usual caveat: none of this matters if you don't run PixlStash. For a plain local setup, ComfyUI's own Load VAE does the same job with a dropdown. This node is for people whose model library already lives in the vault.