boytjj / city-gan

City-GAN is a Generative Adversarial Neural Network that generates typical facades of various cities around the world.

This is the code for the City-GAN paper.

For training run

./gan.py --dataset gan-city --dataroot dataset --batchSize 128 --imageSize 150 --cropSize 128 --hack

If you have a GPU, add the --cuda option for faster training.

You can inspect samples using tensorboard. The following is an example where the rows are: the average of all cities, Amsterdam, Washington D.C., Florence, Las Vegas and Manhattan. The columns are eight randomly chosen inputs to the generator.

It is also possible to transition between cities. The following image has Amsterdam on the left side and Florence on the right one.