What is it about?
WaveStitch is a diffusion-based method for handling various time series related tasks such as imputation, forecasting, and data generation. Importantly, we do conditional (a.k.a constrained) generation, where the outputs have to match the given metadata and any partially observed signal values. The method flexibly guides a pre-trained model at inference-time to adhere with the given conditions. Another key feature is speed. The standard way of handling sequential data like time series or text is autoregression, where generated outputs are fed back as inputs for the next step. For speedup, we introduce a new paradigm for generating such sequential data, drawing inspiration from pipeline parallelism. First, we segment the entire sequence into overlapping time windows and synthesise several windows in parallel, ignoring overlap mismatches. We then "stitch" the inconsistencies through a naively parallelizable inference-time loss, gradually propagating dependencies across the entire sequence like a pipeline.
Featured Image
Photo by ALAN DE LA CRUZ on Unsplash
Why is it important?
WaveStitch shows how diffusion models can handle time series tasks in a fast, non-autoregressive, fully parallel manner while still respecting constraints such as known values or metadata. Its pipeline-style stitching mechanism offers a practical way to generate long sequences efficiently, and this idea is timely since a similar approach could potentially be adapted to long text generation, which is a major challenge in current LLMs.
Perspectives
Although this paper is applied in the context of time series generation, I feel the most important takeaway is for future research to extend the ideas to text generation. The concept of "stitching" is highly relevant and valuable for creating long coherent sentences in a scalable manner.
Aditya Shankar
Technische Universiteit Delft
Read the Original
This page is a summary of: WaveStitch: Flexible and Fast Conditional Time Series Generation With Diffusion Models, Proceedings of the ACM on Management of Data, December 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3769842.
You can read the full text:
Resources
Contributors
The following have contributed to this page
