Forge's "Built In" VAE

This is a blank / identity VAE created to solve a frustrating issue with certain checkpoint models in Forge, reForge and AUTOMATIC1111.

I had to choose a "Base model", but this VAE works for SD 1.4, 1.5, 2.0, 2.1, SDXL, Pony, Lumina, Illustrious and NoobAI. Basically, anything that has "baked" or "built-in" VAE and does not require you to enable it.

Why this exists:

When selecting a checkpoint model from the Checkpoints tab, some models include a line under:

Preferred VAE / Text encoder(s) → Built in

This causes issues like:

• Forge, reForge or A1111 not fully clearing the previous VAE

• A misleading “built-in” VAE being shown when there’s none

• The model continuing to use the last selected VAE, even when the field says “Built in”

The result? Shimmering, JPEG-like artifacts, muddy contrast, and other unwanted effects or even crashes.

What this does:

This .safetensors file is a zeroed identity VAE that:

• Performs no encoding or decoding

• Forces your UI to override the baked-in or ghosted VAE

• Cleans up renders by guaranteeing a neutral, untouched output pipeline

In layman's terms, this VAE does nothing. It is blank, null, empty. It forces your WebUI to switch VAE when switching checkpoints.

How to use it:

1. Drop this into your models/VAE/ folder

2. Select your model designated for "built in VAE" under the Checkpoints tab.

3. Click on the 🛠️ icon to "Edit metadata".

4. Under "Preferred VAE / Text encoder(s)", select "Built in.safetensors", then Save.

How to reproduce the issue this fixes:

If you are not sure why a blank VAE is even needed, here is how to replicate the bug:

1. Load a model that requires a VAE, such as an SDXL or FLUX checkpoint.

2. Go to the 🛠️ Edit Metadata, and under Preferred VAE, select a compatible VAE and Save.

3. Now, switch to a model like WAI-NSFW Illustrious or any other that has VAE already built-in.

4. In the metadata, choose "Built in" for the VAE and Save.

5. Switch between the two models again and look at the top center of the UI where Forge displays the active VAE/Text Encoder.

You will notice that the previous VAE remains active, even though the new model is supposed to use none. This causes artifacts or improper rendering or even a crash until you force an override.

That is exactly where this blank VAE comes in. Drop it in, select it manually, and you are safe.

I hope this helps!

You can find a handful of models with baked in/built-in VAE under "Suggested Resources"! ↴