This project provides a complete tool to use CycleGan, a version of GAN that excels at image-to-image transformation without a need for paired dataset. Here I primarily use it to accomplish image-to-image transformation from google satellite images to pre-1915 Japanese Arts. This has the effect of making the satellite map appear as if they were drawn in the style of pre-1915 Japanese Art.
To generate the following images, I webscraped about 2,500 Japanese art pieces from the Library of Congress (https://www.loc.gov/collections/?fa=subject:art+%26+architecture) and about 1,000 satellite map images from google.
10 ~ 40 epoch seems to generate the best images, with Cycleloss factor of 10 ~ 12.5. Due to VRAM constraint, my batch size was limited to 1.
| Google Satellite | Pre-1920 Japanese Art Style |
|---|---|
 |
 |
 |
 |
 |
 |