ModiFace Presents at ICCV23 Paris

October 6, 2023

We are very excited to be presenting at the International Conference on Computer Vision (ICCV) 2023 in Paris, France! We presented our workshop paper SC2GAN: Rethinking Entanglement by Self-correcting Correlated GAN Space at OOD-CV workshop.

Since 2018, we have been working on Generative AI research on computer vision, with the goal to bring the next generation of VTO experiences. We specifically focus on research that bridges the gap between academic GAN research and real user applications, including addressing bias issues to provide a fair model and speed up the models. In this work, we study the biases in generative adversarial networks (GANs) inherited from training data and propose a novel framework to self-correct such biases on face editing tasks. Compared with state-of-the-art methods, we successfully achieve the best disentanglement of certain correlated attributes pairs (e.g., eyeglasses and age) to provide a fair image editing network by only focusing on the target attribute.

Illustration of our method:

Original latent distribution (left) and self-corrected latent distribution (right). The intensity of the color indicates the density of different sub-regions in the learned latent space. The white arrows show the interpolation directions. The eyeglasses' direction and the age direction are more orthogonal to each other in the self-corrected latent space.