What you see in the image above is the result of the work of an Artificial Intelligence capable of analyzing the lines of a drawing to obtain a realistic face, a photo that seems to come out of the real world.
Transforming simple line drawings into photorealistic facial images is not an easy task, but this AI, called DeepFaceDrawing, manages to produce high-quality images from freehand sketches, even if they are not complete.
What it does is use statistics. It doesn’t require detailed eyelid or lip shapes, it plays with approximation and a huge database of faces and facial components. It takes into account how each facial element works and how it can be put together with another, so that there is no Frankenstein left and something that can be realistically assembled into a single image.
Existing solutions tend to adapt too much to sketches, and that makes it require a very professional drawing up front. They manage to implicitly model the space and synthesize an image of the face in it, to remove restrictions and improve the results.
At the moment they have already tested it with 17,000 sketches, obtaining results that have been mostly Caucasian and South American. This could be the result of the source data or the complexity of the face shapes, there are no further details on why.
This technology will be showcased at the SIGGRAPH virtual conference in July, where we expect them to show the code for the software and some sample of the corresponding application, although commercial software will probably take a few months.
As a utility: in addition to the creation of realistic avatars for the world of videogames, it could be used for the recognition of criminals using the sketches drawn at the police stations.