SemanticStyleGAN
SemanticStyleGAN copied to clipboard
How to improve the mask obtained by domain adaptation on bitmoji dataset
I made fine-tuning on the Bitmoji dataset according to the Domain Adaptation section, and the segmentation mask obtained is still far from the truth. Could you please advise how to get results like the pre-trained model BitMoji-512x512.pt
?
Have you tried to remove the --freeze_local flag for fine-tuning on BitMoji?
Yes. I found that this flag has a large impact on the quality of the generated image and the number of fine-tuning steps. Keeping it the model can get a very good mask. And if it is removed, the image can be generated very well, but the mask is much worse.
Yes. You have to limit the number of fine-tuning steps if you remove this flag.