Your Denoising Implicit Model is a Sub-optimal Ensemble of Denoising Predictions

Ge Kan, Jinhu Lu, Tian Wang, Lei Huang, Baochang Zhang, Aichun Zhu, Hichem Snoussi

Abstract: Denoising diffusion models construct a Markov denoising process to learn the transport from Gaussian noise distribution to the data distribution, however require thousands of denoising steps to achieve the SOTA generative performance. Denoising diffusion implicit models (DDIMs) introduce non-Markovian process to largely reduce the required steps, but its performance degenerates as the sampling steps further reducing. In this work, we show that DDIMs belong to our $\textit{ensemble denoising implicit models}$ which heavily rely on the convex ensemble of obtained denoising predictions. We propose improved DDIM (iDDIM) to demonstrate DDIMs adopt sub-optimal ensemble coefficients. The iDDIM can largely improve on DDIMs, but still deteriorates in the case of a few sampling steps. Thus we further propose $\textit{generalized denoising implicit model}$ (GDIM) that replace the ensemble prediction with a probabilistic inference conditioned on the obtained states. Then a specific instance $t$-GDIM that only depends on the latest state is parameterized by the conditional energy-based model (EBM) and variational sampler. The models are jointly trained with variational maximum likelihood. Extensive experiments show $t$-GDIM can reduces the sampling steps to only 4 and remains comparable generative quality to other generative models.

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.

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: Yes

Please Choose The Closest Area That Your Submission Falls Into: Generative models