EECS 394 Seminar: Steering Diffusion Models for Generative AI, From Multimodal Priors to Test-Time Scaling

Abstract: Diffusion models are advancing generative AI across vision, natural language and science. As data and compute scale, these foundation models learn rich multimodal priors. This talk focuses on leveraging these priors for solving complex downstream tasks using test-time scaling — with guidance and reinforcement learning — covering practical methods, trade-offs and examples.

Short Bio: Morteza Mardani is a principal scientist at NVIDIA Research leading generative AI algorithms, while also serving as a visiting researcher at Stanford University and a Distinguished Industry Speaker for the IEEE Signal Processing Society. He previously held roles as a postdoctoral researcher and research associate at Stanford University and was a visiting scholar at UC Berkeley's RISE Lab. He earned his Ph.D. in electrical engineering from the University of Minnesota. His contributions to statistical and generative learning have been recognized with several honors, including the IEEE Signal Processing Society Young Author Best Paper Award.