Efficient Generative Models for Visual Synthesis

While current visual generative models produce high-quality outputs, they suffer from significant computational costs and latency, limiting their applicability in interactive settings. In this dissertation, we introduce a suite of techniques designed to enhance the efficiency of generative models for image and video synthesis. First, we propose distribution matching distillation, a method that enables the training of one- or few-step visual generators by distilling knowledge from computationally expensive yet highly capable diffusion models. Next, we develop improved distillation techniques that enhance robustness and scalability, culminating in a production-grade few-step image generator. This system is now deployed in widely used software, generating hundreds of millions of images annually. Finally, we extend our approach to video generation by adopting an autoregressive paradigm, significantly reducing latency and enabling fast interactive video generation and world simulation.

Name thesis:degree_name

Doctoral

Department dc:contributor.department

Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science

Grantor dc:publisher

Massachusetts Institute of Technology

Year dc:date.issued

2025

Author dc:creator

• Yin, Tianwei

Advisors dc:contributor.advisor

• Freeman, William T.
• Durand, Frédo

Statement dc:rights

• In Copyright - Educational Use Permitted
• Copyright retained by author(s)

Licence dc:rights.uri

• https://rightsstatements.org/page/InC-EDU/1.0/

Handle dc:identifier.uri

https://hdl.handle.net/1721.1/164029

OAI identifier oai:identifier

oai:dspace.mit.edu:1721.1/164029