Why do most GAN-based models use 256x256 image data as input

[mjz0110]

Hello! Thank you for your fantastic work.After reading your paper, I have some quetions.
Why do most GAN-based models use 256x256 image data as input? What problems might arise from using other image sizes? Can UVCGANv2 be trained with images of size 540x640, for example?

[usert5432]

Hi @mjz0110, Thank you for your interest in our work!

Why do most GAN-based models use 256x256 image data as input?

I am not entirely sure why. I believe it is mostly historical. When you develop a new GAN model, you have to compare it against the predecessors. So you have to use the same image sizes as were used before. Since many early works used image sizes of 256x256, now everybody has to compare against them. But this is only my guess, I do not know why.

What problems might arise from using other image sizes?

I think, the biggest issue is that the architecture of the discriminator (PatchGAN 70x70) is fine-tuned for typical scales of objects in 256x256 images. Using different image sizes may lead to a very disappointing performance. I can't think of any other immediate problems.

Can UVCGANv2 be trained with images of size 540x640, for example?

In principle, as long as the image dimensions are divisible by 32, the default UVCGANv2 configurations can be trained on them.

[mjz0110]

Thanks for your reply! It was very helpful.