SDFlow: Non-Autoregressive Time Series Generation via Flow Matching
Researchers propose SDFlow (Similarity-Driven Flow Matching), a non-autoregressive framework for time series generation that operates entirely in a frozen VQ latent space. Unlike autoregressive models, SDFlow uses a global transport map to avoid exposure bias, enabling parallel sequence generation. Key innovations include a low-rank manifold decomposition with a learned anchor prior to handle high-dimensional VQ token spaces, and discrete supervision integration. The method addresses quality degradation in long-horizon generation by eliminating sequential error accumulation. The paper is available on arXiv under ID 2605.05736.
Key facts
- SDFlow stands for Similarity-Driven Flow Matching.
- It is a non-autoregressive framework for time series generation.
- Operates entirely in a frozen VQ latent space.
- Uses flow matching for parallel sequence generation.
- Eliminates exposure bias via a global transport map.
- Mitigates high-dimensionality with low-rank manifold decomposition and learned anchor prior.
- Incorporates discrete supervision into the process.
- Paper available on arXiv: 2605.05736.
Entities
Institutions
- arXiv