Back to search

Institutional Repository of Vilnius University

Deepfakes naudojant neuroinius tinklus /

Abstract

dc:description

The purpose of this work is to analyze the existing deepfake types, compare different neural network models of the same deepfake creation type in created result verisimilitude, creation speed, creativity freedom and used resources and determine neural network model that performed the best. This work contains 2 different sets of tests – “Face Swap” and moving head model deepfake types. The “Face Swap” model was tested with “DeepFaceLab” and „Reface“, one which has to be retrained for each deepfake and the other one pre-trained on massive dataset to minimize the training needed for creation. The moving head model was tested with „First Order Motion Model for Image Animation“ and „Reface“, one using movement tracing and the other using face tracing. The outcome of tests were as follow. The „Face Swap“ method was best performed by a pre-trained model, „Reface“, as it out-bested „DeepFaceLab“ in all tested aspects other than creative freedom. The moving head model method was best performed by a face tracking using model “Reface”, but not by a lot. Both networks work brilliantly from creation speed and amount of used resources point of view, but fall short on verisimilitude, for different reasons. While the movement tracing method is great for tracking the movement of whole body, it loses accuracy in creating good images, especially in big rotation situations where it needs to understand object dynamics and the face tracing method falls short for only being able to do picture alterations on head related objects. The current best option for head model deepfakes seems to be a mix of both tested types.

Degree

thesis:*
Grantor dc:publisher
Institutional Repository of Vilnius University
Year dc:date
2021

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Šmigelskytė, Rasa,
Contributors dc:contributor
  • Raudys, Aistis

Rights

dc:rights
Statement dc:rights
  • info:eu-repo/semantics/openAccess
Language dc:language
lit

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
Vilnius University
Base URL
epublications.vu.lt/oai
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Šmigelskytė, Rasa,. Deepfakes naudojant neuroinius tinklus /. Institutional Repository of Vilnius University, 2021. https://repository.vu.lt/VU:ELABAETD107116744&prefLang=en_US