ModiFace Presents at CVPR 2023 Conference

June 23, 2023

ModiFace is very excited to be presenting at IEEE / CVF Computer Vision and Pattern Recognition Conference (CVPR) 2023 in Vancouver. We presented our poster paper Sparsifiner: Learning Sparse Instance-Dependent Attention for Efficient Vision Transformers [link] and our ECCV paper Exploring Gradient-based Multi-directional Controls in GANs [link] at AI4CC workshop at CVPR.

In addition to creating world-leading virtual try-on in beauty, we also work on fundamental computer vision research which focuses on efficient models as well as the next generation of VTO experiences utilizing generative AI models. Our deep understanding of fundamental research problems helps us to maintain the leading position in our domain.

In the Sparsifiner work, we present a novel technique that accelerates a vision transformer via an attention mechanism to reduce the computation cost. By learning semantic or spatial relations among image patches/tokens, we derive a binary mask to reduce the network computation cost by only performing network calculation on activated neurons.

In the second work, we present a gradient-based method to navigate and understand learned GAN space. Such understanding will enable applications such as face attributes editing, make-up application, aging simulation as well as synthetic data generation. Compared with other state-of-the-art methods, our method achieves the most realistic and accurate editing results. In the future, these works will be applied to our existing products to bring the next generation of VTO experiences.