This AI only needs a one face scribble to create a photorealistic portrait

0
475
This Ai Only Needs A One Face Scribble To Create.jpg
This Ai Only Needs A One Face Scribble To Create.jpg

This AI only needs a one face scribble to create a photorealistic portrait

Artificial intelligence is capable of many things. The same tries to predict the appearance of Tesla’s Cybertruck (with more or less success) that tries to find hidden bombs or recreate Pac-Man from scratch. The possibilities are virtually limitless. There are already AI systems capable of recreating people’s faces that do not exist, but Chinese researchers have developed a system that goes a little further: recreate faces from a doodle.

This technology has received the name of DeepFaceDrawing (PDF) and its operation is most interesting. There are already technologies capable of recreating portraits, such as Pix2pix or sketch fill, but the results are quite poor. DeepFaceDrawing produces realistic images which, in addition, are generated in real-time and are modified as details are added to the sketch.

From a stroke to a photo of someone who doesn’t exist

Scheme

This deep learning system consists of three main modules: CE (Component Embedding), FM (Feature Mapping) and IS (Image Synthesis). CE adopts an auto-encoder architecture and learns separately the broad features of a face: the left eye, the right eye, the nose, the mouth and “the rest”. FM and IS create a deep learning subnet that converts vectors into images. In other words, the network is made up of two subnets.

The CE module converts the sketches of the components (nose, eyes, chin) into vectors with semantically significant characteristics. FM converts these vectors into maps of individual characteristics that are combined according to the structure of the face (eyes up, nose centred, face down …) and, finally, they pass to the IS module, which synthesizes the. With this AI, at least on paper, it would be as simple as drawing the freehand gesture.

SEE ALSO  WhatsApp for Android now has its own tool to create stickers, and it has already arrived in beta

They also ensure that your system is specially designed to create faces based on the fixed structure of a face, but adapting it to other categories (such as everyday objects) is a “challenging problem”. Overcoming distances, there are already technologies that allow it to do something similar, such as the NVIDIA system that is capable of generating photorealistic landscapes from a sketch.