Improving improved photorealism

Developers from Intel Labs use a convolutional neural network to improve synthetic images, increase their stability and realism.

GTA V to Cityscapes

The modifications made by our method are geometrically and semantically consistent with the original images.







The modifications are also stable over time:







Greening burnt grass and hills in California in GTA:







Add reflections to windows and increase the Fresnel effect (for example, on the roof of cars):







Restore roads:

Transfer of GTA V to Mapillary Vistas

Images from this dataset are recorded around the world using a wide variety of cameras. The images are brighter and with higher resolution.







We remove the distant haze and rebuild the road: We







make the grass more voluminous and green:

• Data & Code: github.com/intel-isl/PhotorealismEnhancement
• Article: Enhancing photorealism enhancement