Artificial intelligence is able to convert outline in seconds.
From March 18 to March 21, the GTC 2019 conference is held in California, organized by NVIDIA. On it, representatives of the American manufacturer of graphic processors demonstrated artificial intelligence, which in seconds processes a schematic sketch and turns it into a photorealistic image.
Neuraleta is called Gaugan, which she received in honor of the painter of the Gaugaen field, whose work refers to the direction of postmingnessismism. It is a generative and contest (GAN). It includes two networks, which are called the generator and discriminator. The generator creates an image, and a discriminator, trained in real photos, suggests how it needs to be corrected.
The application interface consists of three tools: brushes, pencil and fill. In addition to them, various textures are available – from the sky to stones and trees. It is enough to choose the desired texture, set the shape, fill, and neural network will create a realistic object in its place.
For training artificial intelligence (AI), one million photos were used, most of which were taken from the portal Flickr. Vice President of Applied Studies in the field of Bryan Catanzaro (Bryan Catanzaro) assured that they observed copyright.
The creators note that the neural car can change the time of day and seasons. For example, if you specify the winter, then not only the landscape will change, which now covers the snow, but also the sky – it will become cloud and gray. The same applies to the trees that throw off the leaves. At the same time, Gaugan understands when objects are faced with water stroke, and reflects them there. Algorithm Multimodal: Even on the two identical sketches, he will create two different images.
During the demonstration, the program made all the processes in seconds in real time. Brian Katanzaro said that such results were due to the Titan graphic processor. However, according to him, with some modifications, the algorithm will be able to work almost on any platform, but the conversion process will take more time.
While Gaugan works only with nature objects, but the authors do not exclude the ability to train the algorithm for making buildings, roads, people and other elements of reality. According to Katanzaro, this tool will be useful to landscape designers, architects and all those related to the creation of virtual worlds. He hopes that the demo version will soon appear on the site NVIDIA – AI Playground, which company also introduced on GTC 2019.
You can pass our test that will check whether you can determine where the generated portrait, and where is the photo of a real person.