BriefGPT.xyz
Jan, 2024
关于Wasserstein距离中扩散模型的一般概率流ODE的收敛性分析
Convergence Analysis for General Probability Flow ODEs of Diffusion Models in Wasserstein Distances
HTML
PDF
Xuefeng Gao, Lingjiong Zhu
TL;DR
使用概率流常微分方程进行基于得分的生成建模已经在各种应用领域取得了显著的成功。本文首次提供了关于概率流常微分方程采样器的非渐近收敛性分析，假定得分估计准确，并在2-Wasserstein距离下建立了一系列ODE采样器的迭代复杂性结果。
Abstract
score-based generative modeling
with
probability flow ordinary differential equations
(ODEs) has achieved remarkable success in a variety of applications. While various fast ODE-based samplers have been proposed
→