Face loss for face reconstruction

[kl2004]

Thanks for sharing the code!

I tried the checkpoint IN256_Base on face images, but the faces in the reconstructed images are not great. I wonder how the reconstructed faces could be improved. Do you plan to include face loss in the total loss?

[vinyesm]

Same observation, here an example with imagenet_256_L. But I think it could be related to training data

[RobertLuo1]

Hi, there, since the face image is hard to collect and the current scale of face images is limited. Considering the downsampling rate which also causes difficulty in accurately reconstructing the small detailed object such as face, we currently seek for more feasible way to solve that problem.
By the way, we observe that including more training data will benefit which can be shown in our pretrain tokenizer version.