williamyang1991
commented
2 years ago Sorry, I made a mistake in the code when printing the progress information https://github.com/williamyang1991/DualStyleGAN/blob/7a1c560f7fefece00a53e4406d624ad4fd8fc742/destylize.py#L192
The information should be printed as
f"[{ii:03d}/{len(files):03d}]"rather than the original
f"[{ii * args.batch:03d}/{len(files):03d}]"So your progress at [448/283] should be [112/283]
I think your Simpsons style is too abstract and different from real human faces, and it is not well aligned (the eye positions of your two style images differ a lot from each other and from the FFHQ faces). These two points make the destylization fail.
Our method is to simulate the fine-tuning behavior of StyleGAN, and style transfer by fine-tuning StyleGAN relies on the close semantic correspondences between the source domain and the target domain, which is true for real faces and cartoon faces. However, in your case, the domain gap is so large that I'm afraid even fine-tuning StyleGAN cannot work, let alone our method. I think you can check the performance of your finetune StyleGAN on your Simpsons styles and see whether it works. If it does not work, I guess our method might not be able to make it, either.
vltmedia
I am experimenting DualStyleGAN with my custom dataset. I have 283 images and I trained StyleGAN for 600 iterations. Then, I started to destylize it. I set 300 iterations and 4 batch. I expected for it to end when it became 283/283 but now it continues. [448/283] Lperc: 1.254; Lnoise: 0.000; LID: 0.100; Lreg: 0.315; lr: 0.000: 100% Also, I'm not sure if the destylization outputs are as they should be.
