Researchers gave an AI some basic information and had the program paint in a city's finer details

Researchers Qifeng Chen and Vladlen Koltun were curious if an AI could create images without having anything to start with, similar to a person creating a painting without a model or reference. Though the images were not realistic, the results were surprisingly close.

The two researchers fed the AI only a semantic label, which gave the program some basics to work with, like where the streets, cars and trees should go. This information was not telling the AI what these pieces should look like, but where they should be positioned in the picture. The program took these indicators and transformed them into models. The researchers posted this process in a video with a side by side comparison before revealing the final results.

Chen and Koltun posted a paper about their process online and posted their code to GitHub, where readers can learn about the project’s development.

Photographic Image Synthesis

Researchers Qifeng Chen and Vladlen Koltun were curious if an AI could create images without having anything to start with, similar to a person creating a painting without a model or reference. Though the images were not realistic, the results were surprisingly close.

The two researchers fed the AI only a semantic label, which gave the program some basics to work with, like where the streets, cars and trees should go. This information was not telling the AI what these pieces should look like, but where they should be positioned in the picture. The program took these indicators and transformed them into models. The researchers posted this process in a video with a side by side comparison before revealing the final results.