Facebook has presented TextStyleBrush, a sophisticated artificial intelligence (AI) model capable of imitating any typeface from the text of an image. In addition, this technology, which is still in development, allows you to edit and replace text in images with astonishing accuracy .
Many current AI systems can mimic text, but their uses are limited – like digitizing – and only work well in specific situations. Facebook's TextStyleBrush is flexible enough to understand the visual style of a single word and create new ones from it .
Facebook indicates that the TextStyleBrush works similarly to the style brush tools in word processors. However, it uses a “holistic” training method that allows for accurate results.
Instead of training the AI with specific parameters about typography and monitoring its style, Facebook extracts the text from an image, optimizes its display for recognition and segments each of its attributes.
After this analysis, TextStyleBrush is able to recognize specific style information and provide instructions for creating new words. The image generator, for its part, takes this data and calculates the output text resolution to match the style of the input .
A self-monitoring system controls the work of the text generator and allows high precision results to be obtained. As Facebook explains, this tool is made up of two parts. A previously trained typographic classification network and a text recognition network, also previously trained .
TextStyleBrush could be useful on and off Facebook. For example, it could be used to create fonts based on handwriting, improve image translation, or refine other augmented reality applications.
Facebook, TextStyleBrush and the challenges of deepfakes
AI has become a tool that evolves by leaps and bounds. However, its uncanny ability to mimic faces or handwritten words present a challenge in avoiding fraud.
In that sense. Facebook ensures that it has openly published the details of TextStyleBrush to encourage investigation and dialogue around deepfakes and thus be able to anticipate possible attacks and hoaxes.
“If researchers and AI professionals can get ahead of adversaries in building this technology, we can learn to better detect this new style of deepfakes and build robust systems to combat them.”
