A image-to-image style transfer is a machine learning model designed to change the art style of a image by providing a "source" image and the desired image to change the style. However, a generic style transfer model doesn't produce good results, so a specifically trained model (that uses images of the original work) is required for e.g. imitate the art style of a specific work. Is tranining a I2I style transfer model on images of copyrighted works protected by "fair use" or similar exceptions in law? Additionally is commercial use of such models (trained on copyrighted works) allowed?

For instance, can a such model legally imitate the art style of specific works like "Berserk", "Sailor Moon", "Hades" (video game of Supergiant Games), "Lore Olympus", etc.?

1 Answer 1


Derivative work of the source image

A "trained model" of that isn't new. Functions that imitate specific styles have been around since the late 90s and early 00s, known as Filters. Taking a photo and applying a filter got you for example an impressionistic image easily, playing with the settings even could make it look like specific artists.

Modern AI-filters are just better than normal filters. That does not change the fact, that you make a derivative work of the source image.

Training data are currently unsolved

It is not currently a solved thing if training data is "fair use" or not. Multiple cases are running.

You must log in to answer this question.