D-Flow: Differentiating through Flows for Controlled Generation

Explore a comprehensive conference talk on D-Flow, a framework for controlled generation in AI models. Delve into the intricacies of differentiating through flows to optimize generation outcomes in diffusion and flow-matching models without retraining. Discover how this approach unlocks powerful tools for inverse problems and conditional generation. Learn about the key observation that differentiating through the generation process projects gradient on the data manifold, implicitly incorporating prior knowledge. Examine the framework's validation on linear and non-linear controlled generation problems, including image and audio inverse problems and conditional molecule generation. Follow along as the speaker covers background information, controlled generation concepts, D-Flow methodology, theoretical intuitions, experimental results, and conclusions, concluding with an insightful Q&A session.

- Intro + Background
- Controlled Generation
- D-Flow
- Theoretical Intuition
- Experiments
- Conclusions
- Q+A