NVIDIA has already learned how to fake videos professionally using machine learning methods, and now the company's engineers have taken photos. The new system of artificial intelligence can not only remove unnecessary objects from the frame, but also restore the image, adding the missing details on the "smeared" areas.
How it works?
From left to right: "spoiled" portrait, traditional methods, NVIDIA algorithm and original
The Deep Fill tool resembles the fill function with the content in Photoshop, but the NVIDIA option is a qualitatively new level. According to the authors of the project, analogues usually focus on processing rectangular fragments closer to the center of the picture. The new system analyzes the entire image, and therefore can work with defects of any shape and size, regardless of the distance to the border. To learn AI, the team used more than 55 thousand templates, according to which the photo was deleted random areas.
The weak point of the algorithms is "free" structures with a large distance between the elements. For example, they can not always recreate the shape of a building. But with the bookshelves and the restoration of persons they are doing quite well. And they can even remove a couple of wrinkles:
About practical application of technology in NVIDIA services while speech is not.
Sources: DPReview , arXiv