Research map / Time series / Causal and generative time series

SDFlow: Similarity-Driven Flow Matching for Time Series Generation

NeurIPS 2026 · Time series › Causal and generative time series · approach: Irregular series and imputation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SDFlow, a non-autoregressive flow-matching model over a frozen VQ latent space with a learned low-rank anchor prior and categorical cross-entropy supervision, aiming to remove AR exposure bias for time-series generation.

Problem

Autoregressive token models for VQ-based time-series generation suffer from exposure bias, causing quality degradation on long sequences; continuous flow matching in high-dimensional VQ latent space is also hard to initialize well.

Approach

Combines a frozen similarity-driven VQ-VAE tokenizer with a DiT flow network that predicts categorical distributions over codebook indices (variational flow matching with cross-entropy), initialized from a kernel-smoothed prior over learned low-rank latent anchors, enabling parallel non-AR generation.

Key results

• At L=256 on ETTh, SDFlow reports Context-FID 0.001 vs 0.021 for SDformer-ar (95.2% reduction) and DS 0.005 vs 0.008.
• On short sequences (L=24), SDFlow reports Context-FID 0.001 on Energy vs 0.003 for SDformer-ar and matches DS 0.006 on Energy.
• Inference for 1,024 samples at L=24 takes 0.04-0.05s versus 0.15-0.40s for SDformer-ar, reported as 3.0x-10.0x speedup.

Surprising findings

• Generating in raw data space with flow matching performs poorly (DS 0.453), and generic Gaussian or learned MLP priors in the low-rank scaffold fail (DS 0.218/0.213) while the learned anchor prior reaches 0.006.
• Training the Stage-2 flow on only 10% of encoded latents still yields near-full DS/C-FID on held-out real distributions.

Limitations

• Uses a dataset-specific VQ tokenizer and codebook; no universal shared tokenizer.
• The theoretical transport bound assumes a semi-orthogonal basis, while the method learns V unconstrained.
• Speedup is wall-clock and based on a single small setup; the paper notes each pass uses the same transformer attention pattern.

Experimental setup

Datasets: Sines Stocks ETTh Energy

Benchmarks: Sines Stocks ETTh Energy

Metrics: Discriminative Score Predictive Score Context-FID

Baselines: TimeGAN COT-GAN TimeVAE DiffWave DiffTime Diffusion-TS FlowTS SDformer-m SDformer-ar TimeMAR-L

Builds on: variational flow matching CatFlow SDformer VQ-VAE Diffusion Transformer

Closest papers by meaning

• MissPath-FM: Flow Matching with Structured Priors for Partially Observed Time Series (NeurIPS 2026 · similarity 0.89)
• Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting (ICLR 2025 · similarity 0.88)
• Discrete Flow Matching (NeurIPS 2024 · similarity 0.88)
• PrismFlow: Residual Dynamics for Flow Matching in Time-Series Generation (NeurIPS 2026 · similarity 0.87)
• Blockwise Flow Matching: Improving Flow Matching Models For Efficient High-Quality Generation (NeurIPS 2025 · similarity 0.86)
• Structured Coupling for Flow Matching (NeurIPS 2026 · similarity 0.86)