gDDIM: Generalized denoising diffusion implicit models
Qinsheng Zhang Georgia Institute of Technology、Molei Tao Georgia Institute of Technology、Yongxin Chen Georgia Institute of Technology
Abstract
Our goal is to extend the denoising diffusion implicit model (DDIM) to general diffusion models (DMs) besides isotropic diffusions. Instead of constructing a non-Markov noising process as in the original DDIM, we examine the mechanism of DDIM from a numerical perspective. We discover that the DDIM can be obtained by using some specific approximations of the score when solving the corresponding stochastic differential equation. We present an interpretation of the accelerating effects of DDIM that also explains the advantages of a deterministic sampling scheme over the stochastic one for fast sampling. Building on this insight, we extend DDIM to general DMs, coined generalized DDIM (gDDIM), with a small but delicate modification in parameterizing the score network. We validate gDDIM in two non-isotropic DMs: Blurring diffusion model (BDM) and Critically-damped Langevin diffusion model (CLD). We observe more than 20 times acceleration in BDM. In the CLD, a diffusion model by augmenting the diffusion process with velocity, our algorithm achieves an FID score of 2.26, on CIFAR10, with only 50 number of score function evaluations (NFEs) and an FID score of 2.86 with only 27 NFEs. Project
中文速览
扩散模型（diffusion model）生成质量虽好，但每次采样需要上千步神经网络推理，速度极慢；DDIM 虽能大幅加速各向同性扩散，却无法直接推广到 CLD、BDM 等更复杂的非各向同性扩散模型。本文从数值求解随机微分方程的角度重新审视 DDIM，发现其加速效果来源于对得分函数（score function）的一种特殊近似——在概率流 ODE 的精确解上，噪声预测量保持常数，因此只需一次得分估计便可大幅减少离散化误差，这也解释了为何步数少时确定性采样比随机采样表现更好。基于这一认识，作者对得分网络的参数化方式做了一处精巧但关键的修改，提出广义 DDIM（gDDIM），使其可适用于任意线性漂移的扩散模型。实验表明，gDDIM 在 BDM 上实现了逾 20 倍加速，在 CLD 上仅用 50 步便在 CIFAR-10 上取得 FID 2.26、仅用 27 步取得 FID 2.86 的成绩，为非各向同性扩散模型的快速采样提供了一套通用且几乎无额外训练成本的解决方案。
原文 arXiv:2206.05564;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2206.05564v2