Improving Denoising Diffusion with Efficient Conditional Entropy Reduction
Keywords: Denoising Diffusion, Conditional Entropy
Abstract: Diffusion models (DMs) have achieved significant success in generative modeling, but their iterative denoising process is computationally expensive. Training-free samplers, such as DPM-Solver, accelerate this process through gradient estimation-based numerical iterations. However, the mechanisms behind this acceleration remain insufficiently understood. In this paper, we demonstrate gradient estimation-based iterations enhance the denoising process by effectively \emph{\textbf{r}educing the conditional \textbf{e}ntropy} of reverse transition distribution. Building on this analysis, we introduce streamlined denoising iterations for DMs that optimize conditional entropy in score-integral estimation to improve the denoising iterations. Experiments on benchmark pre-trained models validate our theoretical insights, demonstrating that numerical iterations based on conditional entropy reduction improve the reverse denoising diffusion process of DMs. The code will be available.
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2183
Loading