The neural network learned to make sketches, photorealistic landscapes

The development of neural networks is a very fast pace.

Not so long ago neural network for the considered to fake a video, to add to the photos of non-existent details and even generate a person non-existent people.

Now the neural network from Nvidia can turn primitive sketches to photorealistic landscapes. For the synthesis of photorealistic images given the input semantic layout uses the principle of spatial-adaptive normalization. In SPADE affine layer is removed from the map semantic segmentation. In this case, the affine parameters of a spatial-adaptive, i.e. the use of different scaling and offset for each semantic label.

Thus, a neural network, trained on 40 images from thousands of Flickr is the ability to generate your own image based on the vector thumbnails, using preset symbols (blue - sky, green - grass, etc.).

Translated by «Yandex.Translator»