Hi I’m implementing basic artistic style transfer using PyTorch. If I use the features from a late conv layer, the optimized image looks sort of okay (still wrong though) and if I use features from a first conv layer the result is just gray and the loss function does not move at all (in this case the result should be very close to the original image). I suspect the problem might be in pre/post processing but I am not sure.
Notebook documenting the bug
How it should look (fastai notebook)