High-quality and controllable time series generation with diffusion in transformers

ICLR 2025 Conference Submission1629 Authors

Keywords: Diffusion in Transformers, time series generation

Abstract: Current research on time series generation frequently depends on oversimplified data and lenient evaluation methods, making it challenging to apply these models effectively in real-world scenarios. Diffusion in Transformers (DiT) has demonstrated that the traditional inductive biases in neural networks are unnecessary. This paper shows that the advantages of DiT can be extended to time series generation. We add the attention mask and dilated causal convolution to introduce the temporal characteristic. Additionally, we introduce a novel smooth guidance policy for style control during generation, leveraging a property of the diffusion process. Furthermore, our proposed model can generate longer sequences with training in short sequences. Experimental results reveal that our variant of DiT achieves state-of-the-art performance across various data types.

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 1629