报告题目：Score-Based Generative Modeling through Stochastic Differential Equations

报告人：温馨 副教授 长春大学

报告时间：2026年10月27日下午19:00-20:00

报告地点：线上#腾讯会议：834-298-229

校内联系人：邹永魁 邮箱[email protected]

报告摘要：Score-based generative models and denoising diffusion probabilistic models are two prominent classes of generative modeling methods, both of which can be viewed as discrete processes of gradually corrupting data with noise and then reversing the corruption. This paper adopts stochastic differential equations (SDEs) as a unifying perspective: by generalizing discrete time steps to a continuous-time diffusion process, it shows that SMLD and DDPM correspond respectively to discretizations of the variance-exploding (VE) and variance-preserving (VP) SDEs, and further proposes a sub-VP SDE to improve likelihood performance. The key observation is that the reverse-time SDE depends only on the score function of the marginal distribution, which can be estimated by a time-dependent neural network, thereby supporting sampling with arbitrary numerical SDE solvers. On this basis, the paper introduces a predictor-corrector (PC) sampler that significantly improves sample quality under the same computation, and derives the probability flow ODE, which enables exact likelihood computation, adaptive black-box sampling, and uniquely identifiable encoding. Through conditional score adjustment, a single unconditional model suffices for class-conditional generation, image inpainting, and colorization. Experiments achieve FID 2.20, Inception Score 9.89, and a likelihood of 2.99 bits/dim on CIFAR-10, and generate high-fidelity 1024×10241024×1024 CelebA-HQ images from a score-based model for the first time.

报告人简介：

温馨，长春大学数学与统计学院副教授，2024年获色色视频 应用数学博士学位，师从邹永魁教授。主要研究方向为偏微分方程数值解及应用，主要从事基于偏微分方程的遥感影像重建、图像恢复算法研究，曾参与国家重大科研项目、国家自然科学基金项目，在IEEE‑TGRS等国内外期刊发表多篇学术论文。