About the performance of BMSG-GAN

[Iceland-Leo]

Hi, you published BBMSG-GAN. And In order to match the performance of ProGANs, you increased the size of the Discriminator (the one here in BMSG-GAN is technically half that of ProGAN) and use tricks like Batch Spoofing. So, do you mean that the performance of Pro-GANs is better than BMSG-GAN? But BBMSG-GAN is better than Pro-GANs? Of course, the number of parameters is bigger?

[akanimax]

@Iceland-Leo
Sorry for the late reply. I have been lesser and lesser time these days.
Long story short: This repo doesn't use batch-spoofing and the effective discriminator is half of ProGAN's discriminator.
The BBMSG-GAN fixes that and the discriminator is comparable to ProGAN's discriminator. Please note that the Generator is still the same for both.
Btw, the capacity problem was also admitted by styleganv2. So, basically my conjecture was true 😄

Hope this helps 👍.
Cheers 🍻!
@akanimax

[Iceland-Leo]

OK！Thank you so much

[Iceland-Leo]

@akanimax Sorry, one more question. GAN is unstable. May I ask how to get the IS score in your article experient? Is it the average of some epoch results? Or simply choose the maximum IS score?