Particle Guidance: Non-I.I.D. Diverse Sampling with Diffusion Models

Explore a comprehensive conference talk on Particle Guidance, a novel approach to improving diversity and sample efficiency in generative models. Delve into the theoretical foundations and practical applications of this technique, which extends diffusion-based generative sampling by introducing a joint-particle time-evolving potential to enforce diversity. Learn about the analysis of the joint distribution generated by particle guidance, methods for learning optimal diversity-achieving potentials, and connections to other disciplines. Examine empirical results in conditional image generation and molecular conformer generation, where particle guidance demonstrates significant improvements in diversity and accuracy. The talk covers background information, the core concepts of particle guidance, its connections to other methods, fixed and learned potential implementations, experimental results, and concludes with a Q&A session.

- Intro + Background
- Particle Guidance
- Connections
- Fixed Potential Particle Guidance
- Experiments
- Learned Potential Particle Guidance
- Conclusion
- Q+A