Institutional Repository of Vilnius University
GAN modeliai, naudojami internetinėse meninių vaizdų kūrimo svetainėse /
Abstract
dc:descriptionIn the “GAN Models Used in Artistic Imagery Generation Websites” thesis, the generative adversarial network models used in the most popular image generation websites are analysed and one network is chosen. With this network, experiments are carried out in order to ascertain what kind of influence several of the more rarely discussed parameters have on the results of the image generation. This thesis briefly introduces the current situation in image generation and the popularity of websites offering such service. Following this, the architecture, loss and optimization algorithms of generative adversarial networks are described, as well as the challenges that may arise while working with such networks. Furthermore, several popular image generation websites that use GANs are mentioned, and the way the specific GAN architectures they use work is briefly described. Out of the described architectures, the VQGAN architecture is chosen to perform hyperparameter experiments with and the structure of this network is analysed in more detail. The experimental part iss implemented using the Google Colaboratory platform and the Abstract Art Gallery dataset from Kaggle. In this part of the thesis, the VQGAN network is trained using three different values of the number of transformer blocks and the number of codebook entries hyperparameters. The images generated by these trained models are evaluated using Fréchet Inception Distance. The results of the experiments are as follows: decreasing the number of transformer blocks produces lower quality images and increasing the number of codebook entries produces images of a better quality. It is noted during this experiment, that in order to generate images of actual artistic value that could be evaluated by humans, a longer training process with a higher number of epochs and a lower learning rate is required.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Pumputytė, Miglė Morta,
Rights
dc:rights- Statement dc:rights
-
- info:eu-repo/semantics/openAccess
- Language dc:language
- lit
Identifiers
dc:identifier.*- Repository record dc:identifier
- https://repository.vu.lt/VU:ELABAETD192832908&prefLang=en_US