SEAN
SEAN copied to clipboard
ZPdesu
some questions about training details
hello, question1:did you try synthesis of 512*512 size ? Can the model solve synthesis mission of 512 size ?i trained 512 size model with celeba-HQ, synthesis images are not clear and not similar to input images, i just trained 3epoches, question2: did you try IN instead of BN in generator?i found synthesis results of test datasets ,not very clear, lost some details,i guess using IN instead of BN may improve synthesis results.
-
Training on 512*512 is very slow, so I didn't try it. If you wanna improve the training speed, the best way is combing the left-right labels (e.g. left eye and right eye).
-
According to my experiments, IN only performs better during training, and the test performance would be worse than BN (e.g. FID score).
Thank you for your reply. I read your paper,but not very understand how you design your SEAN block . What is your consideration of crossing style map features and segmentation mask features? Can we remove branch after the segmentation mask ,only keep the style map branch?
Yep, it still works, but the performance of quantitative evaluation will drop a bit