I am still testing Neural Style Transfer on Google Colab (part 4): effect of TV loss.
This implementation includes TV (Total Variation) regularization loss. It's important in order to avoid high frequency artifacts in the stylized image. I show the outputs with or without the TV loss term.
Now, a good question would be: Why do these high frequency artifacts show up in the first place? Could be because of the solver used. Here Adam vs LBFGS used in the original paper by Gatys and also by Justin Johnson (best neural style transfer implementation!).
Link to Neural Style Transfer Google Colab:
https://colab.research.google.com/git...