Hi,

thanks for the great work!

I have a question regarding the loss function for training autoencoder.

In the paper, you mentioned We train our autoencoder using a simple reconstruction objective (e.g., l2 loss), but in the code
if mode == "autoenc": criterion = ( torch.nn.MSELoss() if args.mode == 'latent' else LPIPSWithDiscriminator(disc_start=10000000).to(device) )
you trained with the LPIPSWithDiscriminator, since you are using pixel mode. So the loss is not just a L2?