VAE Loader (Searge)

This one's easy to explain: it loads a VAE from a file, the same as ComfyUI's built-in VAELoader, just with Searge branding so it fit neatly into the old Searge workflows. Pick a .safetensors VAE from your ComfyUI/models/vae folder, get a VAE output to feed a decode. That's the node.

It's in Searge/_deprecated_/Files, which tells you the honest recommendation up front: the core VAELoader does the identical job, and the current EVOLVED workflow doesn't need this branded version. It's kept for backward compatibility with older Searge graphs. But the reason you'd load a VAE at all on SDXL is worth getting right, because it's the fix for one of the most common SDXL complaints.

How it works

A VAE is the codec between pixels and the latent space diffusion actually runs in - it's not a colour filter, it's the thing that turns the model's output into a viewable image. Most modern checkpoints bake one in, so usually you don't attach an external VAE at all. The case where you do, on SDXL, is the black-image problem.

The original SDXL VAE generates NaNs in fp16 because its internal activation values overflow, and NaNs decode to solid black. The community fix is sdxl-vae-fp16-fix by madebyollin - a refit that keeps the output identical while scaling the internals to stay inside fp16 range. The Searge README lists exactly this file as optional and says plainly that it "should fix the issue with generating black images." SeargeVAELoader is how an old Searge workflow loads that VAE in.

The inputs and outputs that matter

Minimal, as you'd expect:

• vae_name (required) - the dropdown of VAE files in ComfyUI/models/vae. Pick your file here.
• output VAE - wire it into a decode, or into a Searge sampler that takes a vae input like the img2img samplers.

How to install it

Comes with the Searge SDXL pack. ComfyUI Manager: search SeargeSDXL, install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/SeargeDP/SeargeSDXL.git

then restart. Manual installers need python -m pip install opencv-python run once. To actually have a VAE to load, grab the sdxl-vae-fp16-fix file (the README links it) and drop it in ComfyUI/models/vae; if the folder doesn't exist, create it.

Common issues

The trap is loading a VAE you don't need. Most SDXL checkpoints - and nearly all the popular fine-tunes - already have the right VAE baked in, and attaching a mismatched external one gives you washed-out, greyish colour. The rule: there's no universally "better" VAE, only the one that matches the checkpoint's training. Reach for this node deliberately, in two situations - you're getting black images on fp16 (load sdxl-vae-fp16-fix), or your checkpoint genuinely shipped without a VAE. Otherwise leave it out. And since it's deprecated, prefer core VAELoader in new work; keep this for opening old Searge graphs, and update the pack if it errors after a ComfyUI update.