VAE Loader

This loads a VAE - the codec that turns the latents your sampler produces back into an actual image. It's the same core job as ComfyUI's built-in VAE loader, with one ArtVenture twist: a string override so an external caller can pick the VAE by name instead of you choosing it from the dropdown. Handy for automated pipelines, invisible the rest of the time.

Before you reach for it, the honest question is whether you need a separate VAE at all. On SD 1.5 the answer was usually yes - load the "840k" VAE and the grey, washed-out look goes away. On basically everything since, the answer flipped. Modern checkpoints bake their VAE in, and attaching an external one is now the thing that breaks output, not the thing that fixes it. So the rule that actually matters: the right VAE is the one the checkpoint was trained against. If your model page says the VAE is included, don't load one here.

How it works

It reads a VAE file from your models/vae folder (plus any built-in options the pack exposes) and returns a VAE object your VAE Decode node uses. There's no magic - a VAE is a fixed encoder/decoder pair, not a quality knob you dial. Loading a mismatched one is exactly how you get grey, faded, or noisy output, because you're decoding latents with the wrong codec.

Inputs and outputs

• vae_name (required) - the dropdown of VAE files it finds, plus any built-in entry the pack ships (you'll see options like pixel_space in the list). Pick the one matching your checkpoint, or the one your model explicitly recommends.
• vae_override (optional, string, default None) - name a VAE here to override the dropdown, for automation. Leave it None for manual use.

The output is a single VAE - into your VAE Decode (or VAE Encode, for img2img).

Installing it

Comes with the ArtVenture pack. ComfyUI Manager: search comfyui-art-venture, install, restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/sipherxyz/comfyui-art-venture

Restart ComfyUI. The node is code; VAE files are downloads. For SD 1.5, vae-ft-mse-840000-ema-pruned is the standard. For SDXL, if you're getting black images in fp16, grab the sdxl-vae-fp16-fix. Newer bases generally don't need anything here at all.

Common issues

Grey, washed-out colors. Either no VAE where one is needed (old SD 1.5 checkpoint) or the wrong VAE for the model. Match it to the checkpoint's training. This is an SD 1.5-era symptom; on a 2026 model it almost never has this cause.

Black images on SDXL in fp16. The stock SDXL VAE overflows in half precision. Load the sdxl-vae-fp16-fix instead - it keeps the output identical while keeping the internal math in range. Don't reach for --disable-nan-check; that hides the corruption rather than fixing it.

Output got worse when you added a VAE. Classic mistake on a modern base. Illustrious-family and most current checkpoints bake the VAE in; bolting an external one on degrades the image. Set it back to the model's own and the problem disappears.

Empty or short dropdown. No files in models/vae. Download the VAE your model needs and restart.