Tag
This paper introduces SDFlow, a similarity-driven flow matching framework for time series generation that addresses exposure bias in autoregressive models. It achieves state-of-the-art performance and inference speedups by operating in the frozen VQ latent space with low-rank manifold decomposition.