Analytic-DPM: an Analytic Estimate of the Optimal Reverse Variance in Diffusion Probabilistic Models

• DOI: 10.48550/arxiv.2201.06503
• Publication Date: 2022-01-01
• Authors: Fan Bao; Chongxuan Li; Jun Zhu; Bo Zhang
• Keywords: Divergence (linguistics); Delta method

Diffusion probabilistic models (DPMs) represent a class of powerful generative models. Despite their success, the inference DPMs is expensive since it generally needs to iterate over thousands timesteps. A key problem in estimate variance each timestep reverse process. In this work, we present surprising result that both optimal and corresponding KL divergence DPM have analytic forms w.r.t. its score function. Building upon it, propose Analytic-DPM, training-free framework estimates using Monte Carlo method pretrained score-based model. Further, correct potential bias caused by model, derive lower upper bounds clip for better result. Empirically, our analytic-DPM improves log-likelihood various DPMs, produces high-quality samples, meanwhile enjoys 20x 80x speed up.