With new technology, Facebook wants to stimulate research and a discussion about the use of deep fakes in fonts in images. Facebook has introduced the deep learning system Text Style Brush, which should make it possible to copy any style of text in images. A single word in the picture should suffice for this. Text in images should be able to be edited or completely replaced. According to Facebook, this also works with handwriting.
It is true that replacing text in images is already possible with some other AI systems, as Facebook writes. So far, however, this has been limited to ” clearly defined, specialized tasks”, but developing a generally applicable system for this is much more difficult. After all, it is not only the typography or calligraphy that has to be examined but also other properties of texts and fonts such as transitions, rotations, curvatures, or the changes in the paper that occur when written on with a pen. All of this makes it difficult for automated systems to correctly recognize and reproduce the text in images.
The Facebook system should not only work significantly better than previous technology. This is mainly due to the fact that so far the focus has been on learning the font or target style. With Text Style Brush, Facebook is instead betting on representing the entire appearance of the word in one picture.
Facebook writes about the system:
“By openly publishing these research results, we hope to stimulate additional research and dialogue to prevent deep-fake text attacks similar to deep-fake faces.”
With the help of research, it should be possible to recognize such attacks or manipulations via deep fake text.
But Facebook also hopes to be able to use AI technology productively. For example, text in images could be translated and displayed directly in the same style. Likewise, images could also be more easily personalized or even provided with labels. Facebook hopes that such translations could possibly even be used in AR systems for street signs and the like.