Latent Space

Variational Auto-Encoders are artificial neural networks, usually employed as artistic generative tools. More interestingly, they can compress important features in the drawings. As a result, every element in my database can be encoded with only few dimensions, offering various exploration directions.


Original samples from the database.
Reconstruction of the same samples through the compression in 32 dimensions of a neural network.
Generation of new/never seen elements.

‘Time’ sequences of elements generated along a continuous path in 32 dimensions.

Considering the space of all possible drawings and selecting one pixel among the usual xy dimensions, these small images show orthogonal hidden dimensions.