About

This workshop will focus on key engineering considerations that go into the design of diffusion models in practice. For instance, the practice of generative modeling has surfaced new notions of efficiency and scalability that are yet unexplored in the theoretical literature, including the role of parallelization to accelerate both training and sampling, the importance of the choice of neural architecture for score estimation, and the ability to distill samplers that take many steps into ones that take only a few or even one step to generate an output. In addition, recent developments have pushed the frontiers of diffusion to new data modalities like text and molecules, offering competitive alternatives to traditional autoregressive language modeling and molecular dynamics-based simulation respectively. Finally, the workshop will also touch upon some safety-related aspects of diffusion models, namely watermarking and mitigation of memorization of training data, topics that have already been the subject of fruitful interactions with the TCS community.

Chairs/Organizers