StyleGAN2-Space Navigator Unsupervised Discovery of Semantic Directions for Controllable Face Synthesis
الباحث الأول:
Ali Abdulazeez Mohammed Baqer Qazzaz
الباحثين الآخرين:
م. يوسف سامر المظفر
المجلة:
International Conference on Informatics, Multimedia, Cyber and Information System (ICIMCIS)
تاريخ النشر:
15 ديسمبر، 2025
مختصر البحث:
Generative Adversarial Networks (GANs),
especially StyleGAN2, produce very realistic artificial human
faces, yet control over their output remains a significant
challenge. This paper introduces a technique for exploring the
latent space of a…
Generative Adversarial Networks (GANs),
especially StyleGAN2, produce very realistic artificial human
faces, yet control over their output remains a significant
challenge. This paper introduces a technique for exploring the
latent space of a pre-trained StyleGAN2 model, and it allows for
the semantic control of facial features. The proposed
methodology suggests applying Principal Component Analysis
(PCA) directly to the W+ latent space to perform unsupervised
detection of the most significant variation directions. Both
quantitative and qualitative analyses proved that the major
components align with useful semantic features, including
gender (Component 0), age (Component 5), and pose
(Component 8). The efficiency of the proposed methodology is
presented in StyleGAN2-Space Navigator, an interactive
Gradio-based interface in which users can control these detected
features in real time. The methodology also allows us to generate
a potential semantically controlled synthetic dataset, which is
considered a way to decrease bias in existing face datasets. This
work provides a strong, reproducible method for controlling
generative models, providing a strong way for more
interpretable and steerable AI systems.
K
