[11:53 Mon,25.March 2019 by Thomas Richter] |
The neural network called GauGAN was trained by Nvidia&s researchers with one million Flickr photos of natural landscapes - learning the relationship between more than 180 objects such as snow, meadows, water, stones, hills or mountains. Therefore it is able to recreate landscapes interactively. To do this, the user creates an extremely rough sketch of the desired landscape - where rocks, water, sky, trees or snow should appear in the image - then it generates the desired landscape photorealistically. This can also be edited later by changing its shape or individual elements. ;) Sketch and generate landscapes
By learning the objects and their relationships to each other, GauGAN reacts quasi-intelligent: if a pond is drawn, neighboring elements such as trees and rocks appear as reflections in its water surface. If a grass surface is changed into a snow surface, the whole picture turns into a winter scene, trees that were previously covered with leaves become bare. GANs (Generative Adversarial Networks) can, however, provide visually convincing results due to their structure as a cooperating pair of networks. They consist of a generator and a discriminator. The generator generates images that it presents to the discriminator. Trained on real images, the discriminator gives pixel-by-pixel feedback to the generator on how it can improve the realism of its synthetic images. After training with real images, the discriminator knows that real ponds and lakes contain reflections - so the generator learns to create a convincing imitation. The tool also allows users to add a style filter, i.e. an image can be adapted to the painting style of a particular painter or a day scene can be transformed into a sunset scene. While the GauGAN app focuses on natural elements such as land, sea and sky, the underlying neural network is able to fill other landscape elements such as buildings, streets and people. Nvidia&s newly launched  Playground allows you to try out some of the latest deep learning tools online: not GauGAN yet, but a tool to remove objects from images and soon tools to transfer styles of famous painters to your own photos and generate photorelastic faces.  more infos at bei blogs.nvidia.comdeutsche Version dieser Seite: Photorealistische Landschaften malen per KI |
 |
|
|