instead of using Standard Autoencoder ,try to use Variational Auto encoder , reconstruct images from as low as information, set latent dim 3