Landscape Generator

Title

Conditional Variational Auto-Encoding for Landscape Generation

Abstract

This project concerns the development of a model capable of generating diverse landscape images, including coasts, forests, deserts, glaciers and mountains.

The employed strategy is to design a Conditional Variational Auto-Encoder (CVAE), trained to reconstruct landscape images from the training set with minimal information. The encoding component of the CVAE is trained to extract the most relevant information from the input, learning a transformation from the space of landscape images to an under-dimensioned space, called latent space. The decoding component of the CVAE is trained to learn the inverse transformation, generating landscape images from points in the latent space.

Constraints on the regularity of the latent space enable the application of the decoder to random samples of the latent space, generating new landscape images different from those of the training set. Furthermore, training the CVAE with knowledge about the observed landscape types, that is a condition, allows specifying the desired landscape type during generation.

In conclusion, the results obtained by several configurations of the model have been analyzed, identifying possible improvements and ideas for future explorations, such as designing more complex conditions.