Question about data normalization

[TArdelean]

Hello,
Firstly, thanks for the great research and sharing your implementation!
I have a question regarding data normalization when using the sd-vae-ft-mse latent encoder. In your source code you encode_pixels before applying the VAE, which scales the images in [0, 1]. However, as far as I can tell, the SD VAE was trained using a [-1, 1] normalization (huggingface/diffusers#3726 (comment)).
Could you please clarify this discrepancy?

Thanks,
Timotei

[SonicCodes]

edm2/training/encoders.py

Line 112 in 38d5a70

def encode_pixels(self, x): # raw pixels => raw latents

here's where the encoding happens to be exact , the OpenAI consistency paper apparently discovered this:

@tkarras @TArdelean

Thanks,

[TArdelean]

I see, that clears it up.
Thanks @SonicCodes