Sintetização de rostos humanos por modelos probabilísticos de difusão

Abstract

Human face synthesis is a research area that aims to generate realistic facial images through computational methods. This work explores the use of probabilistic diffusion models for this purpose, relying on an iterative process that starts from a noise-filled image and gradually transforms it into a coherent visual representation. This transformation is guided by a mapping from a low-dimensional latent space to the image space, enabling the model to learn and capture the variability and complexity of human facial features from real data. Throughout the training process, visual samples were generated at different stages, allowing for qualitative observation of the model’s progression. Initially, the generated images resembled pure noise; however, as training advanced, facial contours and details began to emerge, eventually resulting in faces with a high degree of realism. This progression highlights the model’s capacity to reconstruct structured and plausible facial features over time. The results demonstrate that diffusion models are an effective approach for synthetic face generation, showing both training stability and perceptual quality in the produced samples. As a contribution, this study offers a visual analysis of the generative process, illustrating the continuous improvement of image quality during training and reinforcing the potential of diffusion models in facial synthesis applications.