Gaussian Mixture Models Improve DDIM Sampling Quality

A new approach enhances the speed of sampling from pre-trained Denoising Diffusion Probabilistic Models (DDPM) by using a Gaussian Mixture Model (GMM) as the reverse transition kernel within the Denoising Diffusion Implicit Models (DDIM) setup. This technique ensures that the first and second order central moments of DDPM forward marginals match by applying constraints to the GMM parameters. When tested on various models, including unconditional ones from CelebAHQ and FFHQ, as well as class-conditional models from ImageNet and text-to-image generation with Stable Diffusion v2.1 on COYO700M, the results showed that moment matching can produce samples that are just as good or better than those from the original DDIM with Gaussian kernels, especially when fewer sampling steps are taken.