该文提出了一种新的快速ODE求解器DPM-Solver-v3,旨在最小化ODE解的一阶离散化误差。该求解器基于最优参数化,并引入了几个在预训练模型上有效计算的系数,称为“经验模型统计”。此外,还采用了多步方法和预测-校正框架,以及一些技术来提高在少量函数评估(NFE)或大指导尺度下的样本质量。实验表明,DPM-Solver-v3在...
To use DPM-Solver-v3, one can follow the steps below. Special thanks toDPM-Solverfor their unified model wrapper to support various diffusion models. Thenoise scheduleαt,σtdefines the forward transition kernel from time0to timet: $$ p(x_t|x_0)=\mathcal N(x_t;\alpha_tx_0,\sigma...
Official code for "DPM-Solver-v3: Improved Diffusion ODE Solver with Empirical Model Statistics" (NeurIPS 2023) - DPM-Solver-v3/codebases/edm/sample.sh at main · thu-ml/DPM-Solver-v3
Official code for "DPM-Solver-v3: Improved Diffusion ODE Solver with Empirical Model Statistics" (NeurIPS 2023) - History for codebases/edm/sample.sh - thu-ml/DPM-Solver-v3