Institutional Repository of Vilnius University
Difuzijos modelių pritaikymas objektų vaizdų transformavimui į pikselinį stilių /
Abstract
dc:descriptionImage-to-image models can be used to generate desired images by providing them with conditioning photos or textual descriptions that describe the desired output image. One of them is the Stable Diffusion model, but image generation or style transformation using it is generic, limited to the classes the model was trained with. Stable Diffusion can be fine-tuned to generate images of specific objects or types and this work aims to improve pixel art produced by the model. For this purpose, it was adapted using Dreambooth, Textual Inversion and Custom Diffusion methods. Images generated by the fine-tuned models were evaluated by calculating and comparing FID and SSIM metrics. This was done to find out which model's images were most similar to pixel art created by actual artists. A survey was also conducted to find out which fine-tuning method results are most visually appealing to people by transforming the original photo to pixel art. After comparing the numerical metrics, it was found that the images generated by the model fine-tuned with Dreambooth are the most similar to pixel art, but during the survey respondents stated that the pixel style is visually better represented by models fine-tuned using Textual Inversion and Custom Diffusion. They were also rated better at generating pixel art than the not fine-tuned model, while pixel art created by the Dreambooth model was not significantly improved.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2023
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Jasaitytė, Agnė,
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:ELABAETD192846483&prefLang=en_US