Draw a few lines on top of a photo of a face 鈥� the shape of an eyebrow, say 鈥揳nd an AI can turn your sketch into a realistic edit, with no artistic skills required.
Created by Youngjoo Jo and Jongyoul Park at the Electronics and Telecommunications Research Institute in South Korea, the face-editing program is more advanced than simple retouching apps 鈥� it lets you change hairstyles, add smiles, and even insert earrings. It can also convincingly complete a picture of a face that is partially obscured, editing out objects like sunglasses.
The pair used 29,000 celebrity photos to train the face-editing program. Masks were randomly added over the top of these images, training the program to recognise a particular feature 鈥� eyes, for example. Similar masks taught it to recognise the colours and shadows of the face.
Advertisement
The program is a generative adversarial network (GAN), formed of two competing AIs: a generator and discriminator. The generator edits a photo based on an overlaid sketch. The discriminator is given both edited pictures as well as untouched photos, and must distinguish real from fake. After several rounds back and forth, the generator gets so good at producing realistic edits that the discriminator can no longer tell the difference.
鈥淲e think this program would help designers by reducing tedious labour,鈥� says Jo, allowing them to focus on more creative tasks. But, he adds, it can be easily used without any design expertise.
The program is a significant step up in quality from previous face-editing GANs, says machine learning researcher Alex Champandard. 鈥淲hen faced with these kinds of technological innovations, there鈥檚 certainly a lot of both fear and excitement,鈥� he says.
Programs like these will inevitably change the day-to-day work of graphics designers, but he doesn鈥檛 believe they will endanger jobs, says聽Champandard. 鈥淭he big question is, how do we make these systems so that they benefit a lot of the people whose work is being disrupted?鈥�
Reference:
