PFGM++: Unlocking the Potential of Physics-Inspired Generative Models

Explore the cutting-edge advancements in physics-inspired generative models through this 52-minute conference talk by Yilun Xu from Valence Labs. Delve into the innovative PFGM++ model family, which unifies diffusion models and Poisson Flow Generative Models. Discover how these models create generative trajectories for N-dimensional data by embedding paths in N+D dimensional space. Learn about the flexibility of choosing D and its impact on robustness and rigidity. Examine the unbiased perturbation-based objective and the direct alignment method for transferring hyperparameters. Analyze experimental results showcasing superior performance on CIFAR-10 and FFHQ 64x64 datasets, with state-of-the-art FID scores. Gain insights into the improved robustness of models with smaller D against modeling errors. The talk covers an introduction to Poisson Flow Generative Models, experiments on generation quality and speed, the next generation PFGM++, and concludes with a summary and Q&A session.