Winformer: Transcending pairwise similarity for time-series generation

ICLR 2026 Conference Submission14819 Authors

Keywords: time series generation; diffusion model; frequency modeling

TL;DR: We propose a new generative model for time-series with window-wise alignment.

Abstract: Time-series generation plays a critical role in data imputation, feature augmentation, domain adaptation, and foundation modeling. However, the cross-domain generation remains a persistent challenge, as existing methods model time-series interactions either at the granularity of individual points or fragmented segments. This limits their ability to capture and adapt to complex periodic patterns inherent in diverse domains. Specifically, point-wise attention struggles with long-range dependencies, while standard patch-based approaches may break important cyclical structures. To address this, we introduce Winformer, a novel diffusion model framework built on a window-wise Transformer. We shift the fundamental processing unit in the attention mechanism from pairwise points similarity to continuous windows comparison of the entire horizon. By operating on semantically richer window representations, the proposed approach effectively learns and transfers complex periodic patterns across domains. Extensive experiments on 12 real-world datasets demonstrate Winformer's effectiveness, achieving an average performance gain of 10.67% over SOTA baselines.

Primary Area: learning on time series and dynamical systems

Submission Number: 14819