Arbitrary Style Transfer in Real-Time with Adaptive Instance Normalization
Introduces adaptive instance normalization (AdaIN), enabling arbitrary neural style transfer in real time with a single feed-forward network.
Gatys et al.'s neural style transfer renders a content image in another's style but relies on slow iterative optimization. Fast feed-forward methods speed this up but are tied to fixed styles. The paper presents a simple approach that for the first time enables arbitrary style transfer in real time. Its core is an adaptive instance normalization (AdaIN) layer that aligns content feature mean and variance to style features. It matches the fastest method's speed without predefined styles, and allows content-style trade-off, interpolation, and spatial control from one network.
Based on: Arbitrary Style Transfer in Real-Time with Adaptive Instance Normalization · IEEE International Conference on Computer Vision
Curated by Aramai Editorial
Read summary →