Tencent engineers have created an artificial intelligence system that is designed to improve the quality and repair damage in old photos. Based on data from two IE models, it fills the missing pieces of photos in a matter of seconds, offering quite high accuracy and quality.
Usually, such projects are a set of thin settings for existing IE models that are trained in the differences between artificial and real images. Tencent specialists say that this approach often results in poor results. Alternatively, they trained the NVIDIA StyleGAN-2 model in several standard stages of the image creation process, thus keeping a person's "identity" in the original photograph, focusing on the main features of the face, such as eyes and mouths.
The authors of the project have decided to make the results available to anyone who wants to do it. You can either try the algorithm's work on its demonstration, or use the system code if you want to apply it to your project. The project is still far from perfect, but now its assumptions about missing images are quite accurate. Researchers have warned that some changes in the face of real people are possible, and with very high permits, the system cannot yet operate.