Stylizing input images

One or more non-transitory computer-readable storage media storing instructions that, when executed by one or more computers, cause the one or more computers to perform operations for processing an input image to generate a stylized image using a style transfer neural network, the style transfer neural network comprising a conditional instance normalization layer between a first neural network layer and a second neural network layer, the operations comprising:

• maintaining data that maps each image style in a set of image styles to respective parameter values of the image style;
• receiving an input comprising an input image and input style data, the input style data identifying, from the set of image styles, one or more image styles to be applied to the input image to generate a stylized image that is in an input style;
• mapping, using the maintained data, each of the one or more image styles identified by the input style data to respective parameter values;
• determining parameter values for the input style based on the respective parameter values mapped to the one or more image styles identified by the input style data;
• setting current values of parameters of the conditional instance normalization layer to be the determined parameter values for the input style without modifying current values of parameters of the first neural network layer and the second neural network layer;
• processing the input image, using the style transfer neural network having the current values of parameters of the conditional instance normalization layer set to be the determined parameter values for the input style and the current values of parameters of the first neural network layer and the second neural network layer unchanged, to generate the stylized image.

A system implemented by one or more computers, the system comprising:

• a style transfer neural network that is configured to process an input image to generate a stylized image from the input image,
• wherein the style transfer neural network comprises a conditional instance normalization layer between a first neural network layer and a second neural network layer;
• and
• a subsystem configured to perform operations comprising:
• maintaining data that maps each image style in a set of image styles to respective parameter values of the image style;
• receiving an input comprising an input image and input style data, the input style data identifying, from the set of image styles, one or more image styles to be applied to the input image to generate a stylized image that is in an input style;
• mapping, using the maintained data, each of the one or more image styles identified by the input style data to respective parameter values;
• determining parameter values for the input style based on the respective parameter values mapped to the one or more image styles identified by the input style data for the identified at least one image style;
• setting the current values of the conditional instance normalization layer parameters to be the determined parameter values for the input style without modifying values of parameters of the first neural network layer and the second neural network layer;
• and
• processing the input image using the style transfer neural network having the current values of parameters of the conditional instance normalization layer set to be the determined parameter values for the input style and the current values of parameters of the first neural network layer and the second neural network layer unchanged.