Construção de modelos neurais para criação de arte generativa visual

Abstract

The emergence of the generative adversarial networks (GANs) in 2014 allowed the development of neural models that effectively solve computer vision problems such as generation of photorealistic images, image-image translation and super-resolution. Several computer artists have been creatively exploring these generative models and presenting very different results, mainly in the visual field. However, the potential of applying these models in generative art is hampered by problems such as incapacity of element-based generation, low resolution and predictability of results. In the present work are presented two adversarial models, AGAN and bezierGAN that partially counter these problems. The AGAN is an adversarial model with two discriminators, which allows the generation of images with different distributions of the training images, reducing the degree of predictability of the results. The results obtained show that the images generated by the model are abstractions of the training images, and the intensity of these abstractions is controllable. The bezierGAN model is an adversarial generative model for the generation of vectorized images based on bézier curves, which capture the distributions of training bitmap images. The results show that the model can only capture simpler image distributions, failing in complex images. However the great potential of the model is the creation of a latent space for images based on elements, allowing the creation of videos and minimalistic images with organic distributions in any resolution.