This paper introduces a novel method for real-time arbitrary style transfer using adaptive instance normalization (AdaIN). The key idea is to adaptively align the mean and variance of content features with those of style features using a new layer called AdaIN. This approach enables real-time style transfer without being restricted to a predefined set of styles, achieving speed comparable to the fastest existing feed-forward methods. The method is inspired by instance normalization (IN), which has been effective in feed-forward style transfer. AdaIN extends IN by adaptively computing affine parameters from the style input, allowing for flexible user controls such as content-style trade-off, style interpolation, and color/spatial controls. The method is evaluated on various datasets and shows competitive performance with existing methods, achieving significant speed improvements while maintaining flexibility and quality. The approach is also shown to be effective in transferring different regions of the content image to different styles, and allows for real-time adjustments during runtime. The paper also discusses the theoretical implications of the method, highlighting its potential for improving deep image representations.This paper introduces a novel method for real-time arbitrary style transfer using adaptive instance normalization (AdaIN). The key idea is to adaptively align the mean and variance of content features with those of style features using a new layer called AdaIN. This approach enables real-time style transfer without being restricted to a predefined set of styles, achieving speed comparable to the fastest existing feed-forward methods. The method is inspired by instance normalization (IN), which has been effective in feed-forward style transfer. AdaIN extends IN by adaptively computing affine parameters from the style input, allowing for flexible user controls such as content-style trade-off, style interpolation, and color/spatial controls. The method is evaluated on various datasets and shows competitive performance with existing methods, achieving significant speed improvements while maintaining flexibility and quality. The approach is also shown to be effective in transferring different regions of the content image to different styles, and allows for real-time adjustments during runtime. The paper also discusses the theoretical implications of the method, highlighting its potential for improving deep image representations.