The Superposition of Diffusion Models Using the Itô Density Estimator

Learn about a groundbreaking research presentation that introduces SuperDiff, a novel framework for combining multiple pre-trained diffusion models without the need for extensive retraining. Explore how this hour-long talk delves into the theoretical foundations of model superposition, derived from the continuity equation, and introduces a scalable Itô density estimator for calculating log likelihood in diffusion SDEs. Discover how SuperDiff enables efficient model combination through automated re-weighting schemes during inference, mimicking logical operators while maintaining computational efficiency. See practical applications demonstrated across various domains, including diverse image generation on CIFAR-10, enhanced prompt-conditioned image editing with Stable Diffusion, and improved protein structure design. Gain insights into this innovative approach that addresses the growing need to leverage multiple pre-trained diffusion models effectively in the expanding AI landscape.