FastDPM
On Fast Sampling of Diffusion Probabilistic Models (ICML 2021 Workshop)
contribution
- 针对扩散模型提出无需重新训练的快速采样框架,加快推理过程
- 提供多种采样方式,在图像、语音生成等领域进行测试分析
Method
经典扩散模型在反向去噪需要很长的迭代过程( in image synthesis (Ho et al., 2020)),FastDPM缩短采样数()
- 将离散的diffusion step推广到连续diffusion step
- 设计一个建立在diffusion step(t)和 noise level(r) 的双向映射
Bijective mapping between Continuous Diffusion Steps and Noise Level
已知:
定义:
假设是线性分布
则:
给定noise level计算diffusion step
Approximate the Diffusion Process
we aim to construct each step in the approximate diffusion process
noise level序列:
Approximate the Reverse Process
Experiments
- dataset: CIFAR-10, CelebA(163k face images of resolution 64 × 64),LSUN-bedroom (∼3M bedroom images of resolution 256 × 256)
- Models: 预训练模型DDPM,