original paper : https://doi.org/10.48550/arXiv.1703.06868
github link : https://github.com/xunhuang1995/AdaIN-style
AdaIN is a normalization technique introduced in 2017 , succeeding Batch Normalization (BN) and Instance Normalization (IN).
The problem with IN was it normalizes the input image to single style with defined parameters.
BN normalizes a batch of samples to be centered around a single style.
AdaIN can adapt to arbitrary styles in the input image, no need of defined params , it will adaptively compute those param from style input.
AdaIN aligns the mean and variance of content feature maps to those of style feature maps.
Let’s understand the data flow in AdaIN.
There are 2 loss functions : Style loss and Content loss
Net Loss is a weighted combination of both losses.
content loss is the euclidean distance between target feature and output image . The target feature is the AdaIN output , because it leads to faster convergence along with optimizing the decoder network
.
It can be applied to any image size since the network is fully convolutional and the network crops image into random sizes while training.
µ - mean
σ - standard deviation