Supplementary Materials: Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization

In Fig. 1 we show the content and style loss in training and test set. The training loss is post-processed with median filtering of window size 2000 and the test loss is computed every 2000 iterations. We did not observe much overfitting: the training and test loss trend are almost the same. We conjecture that the task is so difficult that the network with millions of parameters still underfit. With a more expressive architecture, the results might be further improved.