Comparative analysis of state-of-the-art deep-learning based face editing algorithms

Abstract

[en] Facial attribute transformation, which involves manipulating specific facial features in images and videos, has become a focal point in computer vision and image processing. This project conducts a comprehensive comparative analysis of cutting-edge methodologies, utilizing diverse models to modify latent imagery representations. We assess various state-of-the-art techniques in facial attribute editing through quantitative, qualitative, and efficiency metrics. Our study demonstrates the superior efficacy of an innovative approach using the Multi-Attribute Latent Transformer Model, which adeptly learns and modifies multiple facial attributes simultaneously. This model not only enhances operational efficiency but also maintains the authenticity and integrity of facial identities. Additionally, we investigate how the correlation of attributes in the training images introduces bias in the results. As part of the project, we have developed a user interface that allows for the visual comparison of four models. This application enables users to observe and compare the distinctions and effectiveness of each model side-by-side. In summary, this research advances the field of facial attribute modification by presenting an in-depth comparative study that highlights the strengths and limitations of leading methodologies in face editing, thereby laying the groundwork for future innovations in refined and scalable facial image transformation.