SDFlow: Similarity-Driven Flow Matching for Time Series Generation
SDFlow: Similarity-Driven Flow Matching for Time Series Generation
Authors: Wei Li, Shibo Feng, Pengcheng Wu, Min Wu, Peilin Zhao Date: 2026-05-07 Paper ID: openalex:2605.05736
Summary
SDFlow is a non-autoregressive framework designed to address the exposure bias and error accumulation prevalent in traditional autoregressive time-series generation models. By operating within a frozen VQ latent space, the model utilizes flow matching to enable parallel sequence generation rather than sequential token prediction. Key innovations include a low-rank manifold decomposition for dimensionality reduction and a variational flow-matching formulation to incorporate discrete supervision. The approach demonstrates superior generation quality and faster inference speeds compared to current state-of-the-art autoregressive methods.
Key Contributions
- Eliminates exposure bias in time-series generation by replacing autoregressive token prediction with a non-autoregressive global transport map.
- Reduces latent space dimensionality via low-rank manifold decomposition and a learned anchor prior for more efficient training and inference.
- Achieves state-of-the-art performance on discriminative and fidelity metrics while providing significant inference speedups over autoregressive baselines.
Open Questions & Future Work
Key Concepts
- sdflow-framework: A non-autoregressive time-series generation framework that uses flow matching on frozen VQ latent spaces to eliminate exposure bias.
Archivist Review
I approved the SDFlow framework as it represents a meaningful shift toward non-autoregressive generation in time series, a domain historically dominated by autoregressive VQ approaches. I also approved the open question regarding universal tokenizers, as this reflects a fundamental architectural limitation for scaling generative models beyond single-dataset benchmarks. Minor naming adjustments were made to ensure the vault remains organized and consistent with existing nomenclature.
Approved Concepts
- Similarity-Driven Flow Matching (SDFlow): Moves time-series generation away from autoregressive VQ-VAE token modeling toward a parallelizable non-autoregressive flow matching paradigm, directly mitigating exposure bias.
Approved Open Questions
- Universal Time-Series Tokenization Bottleneck: A universal tokenizer is critical for transitioning from domain-specific generation models to general-purpose time-series foundation models that can leverage cross-domain data scaling.
Rejected Candidates
- [concept] SDFlow (
sdflow) - duplicate_existing: Renamed to sdflow-framework for consistency with naming conventions and to avoid ambiguity with the paper title. - [open_question] Developing Universal Time-Series Tokenizers (
universal-time-series-tokenizer) - duplicate_existing: Renamed to universal-time-series-tokenization-bottleneck for better alignment with existing architectural bottleneck research tracks in the vault.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2605.05736
- paper_id
- 2605.05736
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-generationflow-matchingvector-quantizationlatent-modeling
- architectures
-
- datasets
-
- concept_slugs
- sdflow-framework
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-05-10T06:03:40Z