This is a slightly modified copy of the Flux.1 Schnell VAE. Which is fine, because it was released under an Apache 2.0 license

The original is a KLautoencoder vae, f8, ch16... however, the wrinkle in the thing is that it makes use of a new feature: "shift_factor"

Seems like it's like "scaling_factor", but a constant offset instead of a multiplier?

The point is, while NEWER models with code in diffusers, support this... the old stable diffusion 1.5 does not.

So, this version of the model actually has the shift factor baked in, with a magical unused tensor (which was probably added to the standard for exactly this reason, but under-documented)

THIS means that the flux.1 vae can now be used to replace the vae's of older models, with a little work.

As you might guess, this means that I am working on a "Flux VAE + sd 1.5" model release.

Stay tuned!