Skip to content
← Projects

Style Transfer

Neural style transfer for images: keep the content of the original image and take on the style of another.

The problem

Create a new image that keeps the layout of the content image but has the brushwork, colours and texture of the style image.

How it works

A pretrained convolutional network extracts features. The output image is optimised to match the content features at a deep layer and the style statistics (Gram matrices) across several layers.

Architecture

content image + style image → CNN feature extractor
optimise the output image: content loss + style loss (Gram matrices)