Hello. Thank you for your contribution. Your code helped me a lot understanding SinGAN. I've changed part of your code.

I think using z_rec as a fixed_Zs[0] (which denoted as z* in paper) might be a better choice rather than just normal z. Also it seems more reasonable to use zero tensors as fixed_Zs[1:] rather than normal z since they use 0 as fixed noise in paper. (though there's no problem running your code since they have been just used to purpose of getting the shape of tensors.)

If you think it's reasonable I hope you merge my code. Once again, thanks for your readable, scalable code!