Performance Analysis of Generative Adversarial Networks and Diffusion Models for Face Aging

• Bruno Kemmer USP https://orcid.org/0000-0002-6830-3281
• Rodolfo Simões USP https://orcid.org/0000-0001-6989-1233
• Victor Ivamoto USP https://orcid.org/0000-0001-5463-3645
• Clodoaldo Lima USP https://orcid.org/0000-0003-3381-5348

Resumo

Computational face aging enables predicting a person’s future appearance using algorithms, with the goal that the output age is close to the expected age and that the individual’s characteristics are maintained. In this work, we evaluate the performance of four generative models on facial aging. Two models are based on generative adversarial networks (GANs), HRFAE, and SAM, and the other two are based on diffusion models, Pix2pix-zero and Instruct-pix2pix. The first two were explicitly trained to generate an aged version of the original person, and the others have a zero-shot generation; in other words, they are generic models that perform different tasks, including facial aging. Since diffusion models have been gaining attention because of their diversity and high-quality image generation, comparing their results against models specifically designed for the task using meaningful metrics is essential. Therefore, we compared these models using the FFHQ Aging database and with the metrics: Mean absolute error (MAE) of the predicted age, Fréchet inception distance (FID), and the cosine similarity of the FaceNet’s embeddings.