Improve Temporal Consistency In Diffusion Models through Noise Correlations
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.
Keywords: diffusion model, temporal consistency, sequential data generation
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Diffusion models have emerged as a powerful tool for generating diverse types of data, including sequential data such as audio, video, and motion. As the temporal consistency in sequential data is crucial for maintaining fidelity and realism, this paper introduce the AutoRegressive Temporal diffusion (ARTDiff) approach to address the challenge of temporal consistency in diffusion models. ARTDiff offers a straightforward and efficient solution that requires minimal computational overhead. Our proposed ARTDiff method leverages the inherent autoregressive dependence structure in time by introducing a Gaussian noise distribution whose correlations between time frames have a functional form in terms of time difference. This design explicitly captures the temporal dependencies and enhances the consistency in generated sequences. We evaluate the effectiveness of ARTDiff on audio and motion generation tasks. Experimental results demonstrate that ARTDiff significantly improves the fidelity and realism of generated samples compared to baseline diffusion models. The simplicity and efficiency of ARTDiff make it a practical choice for incorporating temporal consistency in diffusion-based generation 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.
Submission Number: 3405
Loading