We estimate global class covariances at the server with a provably unbiased estimator requiring only local class means from clients, achieving performance competitive or superior to algorithms sharing second-order statistics

We propose Cradle2Cane, a novel two-pass diffusion-based face aging method that balances age accuracy and identity preservation, achieving state-of-the-art performance on the CelebA test dataset with improved efficiency.

We propose the first multimodal foundation model for 3D genomics, integrating Hi-C contact maps and chromatin accessibility to achieve unified semantic representation, outperforming state-of-the-art methods on diverse tasks.