1: \begin{abstract}

2:   It is well-known that 

3:   training of generative adversarial networks (GANs) requires huge iterations 

4:   before the generator's providing good-quality samples. 

5:   Although there are several studies to tackle this problem, 

6:   there is still no universal solution. 

7:   In this paper,

8:   we investigated the effect of sample mixing methods, that is, Mixup, CutMix, and newly proposed Smoothed Regional Mix (SRMix), 

9:   to alleviate this problem. 

10:   The sample-mixing methods are known to enhance the accuracy and robustness in the wide range of classification problems, 

11:   and can naturally be applicable to GANs

12:   because the role of the discriminator can be interpreted as the classification between real and fake samples.

13:   We also proposed a new formalism applying the sample-mixing methods to GANs 

14:   with the saturated losses 

15:   which do not have a clear "label" of real and fake. 

16:   We performed a vast amount of numerical experiments using LSUN and CelebA datasets. 

17:   The results showed that Mixup and SRMix improved the quality of the generated 

18:   images in terms of FID in most cases,

19:   in particular, SRMix showed the best improvement in most cases.

20:   Our analysis indicates that 

21:   the mixed-samples can provide different properties 

22:   from the vanilla fake samples, 

23:   and the mixing pattern strongly affects the decision of the discriminators.

24:   The generated images of Mixup have good high-level feature but 

25:   low-level feature is not so impressible. 

26:   On the other hand,

27:   CutMix showed the opposite tendency. 

28:   Our SRMix showed the middle tendency, that is, showed good high and low level features.

29:   We believe that 

30:   our finding provides a new perspective to accelerate the GANs convergence

31:   and improve the quality of generated samples. 

32: \end{abstract}

33: