Arbitrary Style Transfer in Real-Time with Adaptive Instance Normalization

Building on Gatys et al.'s neural style transfer—which renders a content image in the style of another but requires a slow iterative optimization—and on faster feed-forward approximations that are limited to a fixed set of styles, this paper presents a simple yet effective approach that for the first time enables arbitrary style transfer in real time. At the heart of the method is a novel adaptive instance normalization (AdaIN) layer that aligns the mean and variance of the content image's features with those of the style image's features, effectively transferring style through feature statistics.

The approach achieves speed comparable to the fastest existing methods while removing their restriction to a predefined set of styles, so it can handle arbitrary new styles. It also enables flexible user controls—including content-style trade-off, style interpolation, and color and spatial controls—all using a single feed-forward neural network. AdaIN became a foundational, efficient technique for style transfer and, more broadly, for conditioning generative networks via feature statistics.