Neural network taught to draw the missing details in the photo of people

Projects based on neural networks are not uncommon. They appear every day. Someone sorts cucumbers, someone draws pictures or composes texts of fake news, but someone restores missing details in photographs of people.

The new project, which, incidentally, has already been posted on GutHub , allows you to restore parts that for one reason or another are missing in the photo. By the way, some details may be a “fantasy” of the program itself. For example, this is a hairdo of a bald man or a smile in a photo where she was not.

Project basis- Generative-competitive neural network SC-FEGAN. Networks of this type work in many similar (and not so projects). Usually they consist of two parts. This project is no exception. The first part is Unet-like, an image generator. The second is the SN-pachGAN discriminator. The generator creates images (which is logical), the discriminator cuts off unsuccessful generations and "makes" a decision that should appear in the photo.

The service works simply - the user needs to upload a photo of a person and create new image details. It can be hair, facial expressions, jewelry. If you wish, you can try to remove some details from the photo by changing the color of the hair or eyes along the way. As mentioned above, you can add hair to a bald man, it all looks quite organic.

For just to use the service, you must follow the instructions of the developer. It is not so simple, but nothing super-complicated is required. The authors of the development plan to make it a part of any commercial applications, including mobile programs or web services.