The NVIDIA GauGAN2 neural network, trained on 10 million nature photos, generates realistic images based on a brief description. Then you can add new objects to the image by drawing their sketch by hand.
GauGAN2 implements segmentation mapping, drawing and text-to-image conversion within a single model, which makes it a powerful tool for creating photorealistic art with a combination of words and drawings.
To do this, the generative-adversarial neural network was trained using the NVIDIA Selene supercomputer. The researchers used a neural network that studies the relationship between words and the visual effects they correspond to, such as ”winter“, ”foggy“ or ”rainbow”.
After generating the image, you can create a semantic segmentation map that shows the location of objects in the scene. This scene can be completed with simple sketches, for example, the sky, a tree, a rock or a river.