Modelling and forecasting art movements with CGANs.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Art History

Background:

• Conditional generative adversarial networks (CGANs) typically use discrete labels for conditioning.
• Generating samples from a sequence of continuous distributions is a challenging task.
• Artistic movements evolve over time, presenting a sequential data problem.

Purpose of the Study:

• To propose a novel method for training CGANs conditioned on sequences of continuous latent distributions.
• To apply this method to generate art samples reflecting the evolution of artistic movements.
• To forecast latent distributions of future art movements and generate representative artworks.

Main Methods:

• Developed a new training approach for CGANs to handle sequential continuous latent distributions.
• Applied the method to a dataset of paintings across different art movements.
• Utilized a vector autoregressive (VAR) model to forecast the means of latent distributions for future art movements.

Main Results:

• The proposed methodology successfully generates samples from a sequence of distributions.
• Vector autoregressive modeling accurately predicted the latent distributions of future art movements.
• Experiments showed small mean distances between predicted and plausible current art movements, validating the forecasting accuracy.

Conclusions:

• The novel CGAN training method enables conditioning on continuous latent distribution sequences.
• The integration of VAR models allows for accurate forecasting of art evolution.
• This approach offers a powerful tool for generating novel artworks and understanding artistic trends.