StyleTransfer problem about shapetexturedebiasedtraining

[Jumpat]

Thanks very much for your help. Your method is proved useful, which really helps me a lot!
Wish you all the best~

[LiYingwei]

Thanks for your interest! I think this noise might because the output images are slightly beyond the image space, which causes an overflow issue -- I remember some values of the pixels are larger than 1. The noise you see is probably due to these pixels are overflow, and then truncated by the visualization tool.

Therefore, it is necessary to rescale the value of the image back to [0, 1] before visualization. However, when using these generated images as training data, I don't rescale them back to [0, 1] image space (unlike the visualization tool, the preprocessing functions won't truncate the overflow pixels)

For the performance issue, could you provide more details about what kinds of dataset/model/training strategies are you using? Maybe I can find some other issues.

Thanks,
Yingwei

[Jumpat]

Thanks for your reply!
The dataset I'm using is just a random small subset of imageNet, and the lemon and chimpanzee you used in your paper are found in imageNet with my own hands😂.
So in fact this noise doesn't have to be removed in the real train process. I will try your method and see whether it help. I will give you some feedback after that!