Baby face generation and edition through text-guided diffusion models

Aguilar Velazquez, Marc

Baby face generation and edition through text-guided diffusion models

License

Citation

Aguilar Velazquez, M. Baby face generation and edition through text-guided diffusion models. 2024. handle: http://hdl.handle.net/10230/68217

This citation was generated automatically.

Abstract

The synthesis of photo-realistic human faces has been a trending area of research since the release of deep generative models. During the last years, synthesized images have become almost indistinguishable from real ones to the human eye. Face image synthesis has a wide array of applications, such as advertising, data augmentation or medical imaging. However, generative models are computationally expensive and require large training datasets to produce high quality outputs. Even though multiple datasets containing adult faces are available, baby face datasets are much more difficult to gather due to privacy concerns. Furthermore, generating photo-realistic baby faces is not a well explored topic. This thesis explores transfer-learning techniques to generate photo-realistic baby faces and presents two novel pipelines to edit them. To achieve this, we propose using a pre-trained diffusion model (DM) and performing transfer learning to adapt it to the specific task of generating high-quality baby faces. The DM is conditioned to facial features, ethnicity and expression by a textual prompt. Moreover, we extend the DM with editing capabilities, allowing for the synthesis of multiple instances of a randomly generated identity while varying non-identity attributes. We address two non-identity modifications based on DM, firstly expressions and then pose orientations. To evaluate the performance of the model, real and synthesized baby faces were presented to 149 observers who judged their realism. Results indicate that faces synthesized by our model are almost indistinguishable from real ones. When distinguishing real from AI images synthesized by our model, 61.1% of the participants failed. In conclusion, this thesis demonstrates the efficacy of diffusion models in generating and editing baby faces.