Jinxin Zhou1* | Tianyu Ding2*† | Tianyi Chen2 | Jiachen Jiang1 | Ilya Zharkov2 | Zhihui Zhu1 | Luming Liang2† |
1Ohio State University | 2Microsoft |
Turning the top to the bottom by adding only three lines of code.
We present DREAM, a novel training framework representing Diffusion Rectification and Estimation-Adaptive Models, requiring minimal code changes (just three lines) yet significantly enhancing the alignment of training with sampling in diffusion models. DREAM features two components: diffusion rectification, which adjusts training to reflect the sampling process, and estimation adaptation, which balances perception against distortion. When applied to image super-resolution (SR), DREAM adeptly navigates the tradeoff between minimizing distortion and preserving high image quality. Experiments demonstrate DREAM's superiority over standard diffusion-based SR methods, showing a 2 to 3x faster training convergence and a 10 to 20x reduction in necessary sampling steps to achieve comparable or superior results. We hope DREAM will inspire a rethinking of diffusion model training paradigms.
We propose a novel training framework for diffusion models, enhances alignment between training and sampling, significantly improving image super-resolution efficiency and quality with three lines of code changes.
DREAM enables much faster training convergence.
DREAM allows a significant reduction in necessary sampling steps.
@article{zhou2023dream,
title = {DREAM: Diffusion Rectification and Estimation-Adaptive Models},
author = {Zhou, Jinxin and Ding, Tianyu and Chen, Tianyi and Jiang, Jiachen and Zharkov, Ilya and Zhu, Zhihui and Liang, Luming},
journal = {arXiv preprint arXiv:2312.00210},
year = {2023},
}