Question about the image encoder

[mdciri]

Hello,

it is not clear to me how the encoder is trained. I understood that you do multiple dow-sampling steps and then two a Linear/Dense layers to calculate the mean and the variance vectors.

which loss function do you use and which are its inputs? So, with what do you compare per every step the output mean and variance vectors?
Why did you choose a length of 256? Is it because you train on 256x256 images or because 256 is the number of grey-scale levels?

Thanks in advance,
Best regards,
Marco Domenico Cirillo