Projects Last Projects

A significant challenge in GAN-based human face manipulation is maintaining a consistent identity, as existing methods often cause changes in facial characteristics. This work presents a solution that integrates personalized generative capabilities with precise image manipulation techniques to address this issue. Our method utilizes a personalized deep generative prior to fine-tune a pre-trained face generator using a small set of portrait images (~100) of an individual, creating a local, low-dimensional manifold in the latent space. This enables semantic editing faithful to the individual’s key facial characteristics while using a user-friendly GUI with a point-to-point dragging method. We additionally enhance the accuracy and stability of image manipulation by employing classical methods such as normalizing latent vectors and expanding our manifold by using known editing direction vectors all that while maintaining comparable runtime efficiency.

Please, see project report.

Please, see final presentation.