CollaFuse: Collaborative Diffusion Models for Efficient Image Generation

A new approach called CollaFuse has been introduced for distributed collaborative training of diffusion models, inspired by split learning. The method addresses challenges in generative AI such as data availability, computational requirements, and privacy. Unlike traditional federated learning, which imposes heavy computational burdens on clients, CollaFuse retains data and inexpensive processes locally while outsourcing computationally expensive tasks. This reduces the burden on individual clients, making it suitable for resource-constrained environments. The approach facilitates collaborative training without compromising privacy or efficiency.