Improved DDIM Sampling with Moment Matching Gaussian Mixtures
Researchers have developed a new method to improve the sampling process in Denoising Diffusion Implicit Models (DDIM). Their approach utilizes a Gaussian Mixture Model (GMM) as the reverse transition operator, which matches the first and second-order central moments of the DDPM forward marginals. This technique has demonstrated the ability to generate samples of equal or higher quality compared to the original DDIM, particularly when using a small number of sampling steps. AI
IMPACT Enhances sample generation quality and efficiency for diffusion models, potentially improving downstream applications.