EGEA-DM: Eigenvalue-Guided Explainable and Accelerated Diffusion Model

ICLR 2026 Conference Submission7464 Authors

Keywords: diffusion model

Abstract: Diffusion models have achieved remarkable success in generating high-quality data, yet challenges remain in training convergence, interpretability, and fine-grained controllability. Additionally, the high computational cost of training is often overlooked from a theoretical perspective. To address these limitations, we propose Eigenvalue-Guided Explainable and Accelerated Diffusion Model (EGEA-DM), a novel framework grounded in ergodic theory. EGEA-DM leverages the L-generator’s principal eigenvalue to precisely control the forward diffusion speed, enabling adaptive adjustment of reverse steps during both training and sampling. By modulating the forward process through the L-generator’s coefficients, our method enhances training efficiency, improves generation quality, and provides interpretability and controllability. Extensive experiments validate the effectiveness of EGEA-DM, demonstrating its potential to advance the practical applicability of diffusion models.

Primary Area: generative models

Submission Number: 7464