Trained with regularization, although couldn't get rid of screencap and vignetting effect anyway. VAE, weight decay and custom LR scheduler were used during training process. Default weight seems like working to me. Can generate dataset-nonentrant things pretty well (see showcase post).