Input Perturbation Reduces Exposure Bias in Diffusion Models

Mang Ning, Enver Sangineto, Angelo Porrello, Simone Calderara, Rita Cucchiara

Keywords: Generative Models, Diffusion Model

Abstract: Denoising Diffusion Probabilistic Models (DDPMs) are fast becoming one of the dominant generative methods thanks to their high generation quality and diversity. However, one of the main problems of DDPMs is their large computational cost, which is due to the chain of sampling steps. In this paper, we argue that one of the reasons why DDPMs need a long sampling chain is due to an exposure bias problem, similar to the analogous problem in autoregressive text generation. Specifically, we note that there is a discrepancy between training and testing, since the former is conditioned on the ground truth samples, while the latter is conditioned on the previously generated results. In order to alleviate this problem, we propose a very simple but effective training protocol modification, consisting in perturbing the ground truth samples to simulate the inference time prediction errors. We empirically show that the proposed input perturbation leads to a significant improvement of the sample quality and to smoother sampling chains, with a drastic acceleration of the inference time. For instance, in all the tested benchmarks, we observed an acceleration over a state-of-the-art DDPM of 12.5 times.

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

Community Implementations: [ 1 code implementation](https://www.catalyzex.com/paper/arxiv:2301.11706/code)