The Stolen Moon

(Sadly, the textures of print & paper are eaten alive by this site's compression)

The footage of this film was created with Generative Adversarial Networks (GAN), a form of machine learning.

Stills from 90 pre-Soviet silent films were extracted to compile comprehensive datasets for training. The datasets corresponded to a variety of categories like Landscapes (with people), Landscapes (without people), Interiors, Marked Frames, Intertitles, etc.. The resulting models were then used to build the sequences seen in the film, reflecting how the machine interprets the film stills it was trained on. To give the digital images a more organic texture resembling early cinema, stills of the sequences were printed with a laser printer and scanned.

The text that appears in the subtitles was collaged from intertitles of the films used for training. They aim to capture the topics and atmosphere of pre-Soviet cinema.