Diffusion and flow-based generative models have become powerful tools for high-quality image generation, restoration, and editing. However, their practical use is often limited by the large number of neural function evaluations (NFEs) required during sampling. This talk will first introduce the basic ideas behind diffusion models, including the forward noising process, reverse denoising process, score/velocity parameterizations, and the connection between sampling and numerical solution of ODE/SDE dynamics. The second part of the talk will focus on the problem of accelerating sampling without additional training or distillation. I will present our recent work, CAB: Corrected Adams–Bashforth, which proposes a training-free sampler for both flow and diffusion models. The method first transforms different sampling dynamics into a common rectified coordinate system and then applies a multistep Adams–Bashforth predictor with a correction term computed from past velocity evaluations, incurring no additional neural function evaluations. The method has at least third-order local truncation error and second-order global error, while improving the quality–NFE trade-off in low-step sampling regimes such as 6–20 NFEs. The talk will discuss the motivation, algorithmic design, theoretical properties, and experimental behavior of CAB on pretrained flow and diffusion models, including large-scale text-to-image settings.
Pravin Nair is currently an Assistant Professor in the Department of Electrical Engineering at IIT Madras. Prior to this, he served as Senior Chief Engineer in the AI Video Processing Labs at the Samsung Research Institute, Bengaluru. He earned his PhD from the Department of Electrical Engineering at the Indian Institute of Science (IISc). He also holds an M.Sc. by Research in Electrical Engineering from IISc and a B.Tech. in Electronics and Communication Engineering from the Amrita School of Engineering. His research focuses on addressing longstanding challenges in image and video reconstruction for both offline and real-time settings, with particular emphasis on developing high-quality, stable, real-time algorithms for image and video processing. He has received several prestigious recognitions, including Gold Medal for the best PhD thesis dissertation at IISc Bengaluru, the Samsung Excellence Super Tech Award for the 10–100× Zoom feature in the Samsung Galaxy S24 Ultra, and the Prime Minister Early Career Research Grant from Anusandhan National Research Foundation.