Institutional Repository of Vilnius University
Dangaus skliauto generavimas pasitelkiant neuroninius tinklus /
Abstract
dc:descriptionThis work aims to solve the problem of generating skydomes from data labels using neural networks for building backgrounds of 3D scenes. In order to solve the problem, a dataset of skydomes was created, a cascaded diffusion model was trained to generate high-resolution skydomes, a DDIM generation method was implemented to shorten the generation process, and a conditional generation method was implemented to control the generation of the model. The influence of variance schedules, number of generation steps and cascaded diffusion model architectures on the generation quality is also investigated. The images generated by the trained cascaded diffusion model showed that this type of neural network is capable of generating skydomes from data labels, generating a large variety of images that reflect the characteristics of skydomes, however the model does not always generate content that accurately matches the labels. The Investigation on the effect of the number of generation steps has shown that with the DDIM generation method, the number of steps can be greatly reduced with negligible loss in generation quality, whereas with the DDPM generation method, the quality of the results rapidly deteriorates as the number of generation steps is reduced. The comparison of the variance schedules showed that, contrary to the literature, better quality images are obtained using a linear variance schedule, as the model is not able to correctly generate the images corresponding to the labels using a cosine variance schedule. Experiments with cascaded model architectures have shown that the best results are obtained by combining standard diffusion models into the longest possible chain.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Jasiūnas, Dominykas,
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:ELABAETD192846516&prefLang=en_US