How would imaginary passports look like? Based on real existing passport covers from 200 countries and territories I’ve trained a machine learning model to generate new passport covers based on the patterns of existing covers. This works with a Generative Adversarial Network (GAN) that generate new images and passed a recognition test to be familiar with the original dataset.

Training the machine learning model

The model has trained with 3000 iterations where it has “learned” a bit more each time. For training I used RunwayML, the powerful machine-learning tool for creators. The model is public and can be found here.

Step 500 of 3000

Early learning progress. The model creates and early understanding of colours and patterns but is still far away from passport covers.

Step 800 of 3000

After 800 steps the model makes progress so better understand the location of text and symbols on passport covers.

Step 1500 of 3000

After 1500 iterations the background has become a solid colour and the text and symbols are at the right place. From the distance this already clearly has similarities with passport covers.

Step 3000 of 3000

The finished model. Slight improvements from 1500 iterations. The model clearly can now replicate the visual identity of passport covers, however, the text does not represent real exisitng text.

Walk through latent space

Discover in this video the possible creations of new generated passport covers.