SomeQuestion

[huangtao36]

Thanks for your good work. I train it from sketch with celebA, but I can't get good quality images. I have some questions and hope you can help.

• How long have you trained a model? and How much GPU memory?
• Have you tried StyleGAN2?
• Are there any tricks or is it difficult to train by itself?

[huangzh13]

In fact, I also failed on the CelebA dataset, which I think is caused by RelativisticAverageHinge Loss (I originally thought it was a good enough loss function). But please try Logistic Loss, which I added later (loss function in the original paper). According to my experiments, it was successful on the CelebA dataset.

• The training efficiency is really low now, and may need more optimization. It usually takes more than 2 days to train 128 * 128 resolution images.
• I'm trying to implement stylegan2 with a similar structure and make it compatible with official pre-trained models.

[huangtao36]

ooh, thanks, this is a great tip, I will try it. And I will try this model on the person images (contain the body), hoping to have a good effect.