Institutional Repository of Vilnius University
Generatyvinių modelių kolapso prevencija /
Abstract
dc:descriptionThis thesis examines the phenomenon of generative model collapse and explores potential strategies for its prevention. Model collapse refers to the process in which artificial intelligence systems, through repeated training on their own generated data, gradually loose the ability to reproduce the original diversity of their training distribution. To investigate potential mitigation methods, this work reviews established evaluation metrics from scientific literature and introduces a new approach: involving human users in the data filtering process. This user‐guided filtering simulates how people choose and share images online, a process that has not yet been studied with real individuals in technical evaluations but crucial in real‐world data selection. The study further explores whether human participants, when educated about diversity and representation, can help counteract collapse instead of unintentionally reinforcing it. In doing so, it tests the hypothesis that human behaviour is not fixed or inherently biased toward collapse and that education may help users act as curators of more resilient training data. The experiment demonstrated that education can indeed have an impact. Users contributed to preserving the data diversity of the StyleGAN2‐FFHQ model more effectively than reported in the existing literature. However, a similar decline in image quality was still observed, manifested through the emergence of artefacts in later generations.
Degree
thesis:*- Grantor dc:publisher
- Institutional Repository of Vilnius University
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Virpšaitė, Greta,
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:ELABAETD239032727&prefLang=en_US