Differentiable Augmentation for Data-Efficient GAN Training Supplementary Material Shengyu Zhao IIIS, Tsinghua University and MIT Zhijian Liu MIT Ji Lin MIT Jun-Y an Zhu Adobe and CMU Song Han MIT

Since we find that random cropping leads to a worse IS, we process the raw images with random scaling and center cropping instead. See Figure 3 for a qualitative comparison between BigGAN and BigGAN + DiffAugment. CR-BigGAN [50] reports an FID of 6.66, which is slightly better than ours 6.80 (BigGAN + DiffAugment) with 100% data. CR-BigGAN only achieves an FID of 7.95 with an IS of 82.0, even worse than the baseline BigGAN. Nevertheless, our CIFAR experiments suggest the potential of applying DiffAugment on top of CR.