Exploring Latent Embeddings in Diffusion Models for Face Orientation Conditioning

Facial rotation is a critical task in computer vision with numerous applications across various domains. Traditional facial rotation techniques, such as reconstruction-based and 3D geometry-based approaches, require ground truth data for training, which presents limitations. This thesis proposes two novel techniques for addressing the facial rotation problem without relying on pairs of images. These techniques leverage the latent space of a DDIM trained to reconstruct human faces and condition the generation process to produce rotated faces. A comprehensive overview of facial rotation techniques and Diffusion Models (DMs) is provided, along with the development, implementation, and evaluation of the newly proposed methods. The first method modifies specific input image pixels, while the second technique fits a linear regressor to sample from the latent space. The second method demonstrates better stability and simplicity, effectively producing rotations of up to ±30◦. Future research directions include using unbiased datasets with greater face orientation variation and improving the second method's computational efficiency.