Mirror Diffusion Models for Constrained and Watermarked Generation

Explore the concept of Mirror Diffusion Models (MDM) for constrained and watermarked generation in this comprehensive talk by Guan-Horng Liu from Valence Labs. Delve into the challenges of applying diffusion models to constrained data sets and discover how MDM offers a solution by learning diffusion processes in a dual space constructed from a mirror map. Examine the efficient computation of mirror maps for popular constrained sets like simplices and ℓ2-balls, and understand how MDM outperforms existing methods. Investigate the potential of constrained sets as a mechanism for embedding invisible watermarks in generated data for safety and privacy purposes. Gain insights into the algorithmic opportunities for learning tractable diffusion on complex domains through this in-depth presentation, which includes a paper discussion and Q&A session.