Research map / Time series / Causal and generative time series

PrismFlow: Residual Dynamics for Flow Matching in Time-Series Generation

NeurIPS 2026 · Time series › Causal and generative time series · approach: Diffusion models for time series · 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 PrismFlow, a flow matching method for time-series generation that adds hard-routed Koopman-inspired residual experts trained with a confidence-aware winner-take-all loss to counter estimator-level mode averaging.

Problem

Single global vector-field estimators in flow matching average incompatible local velocities for heterogeneous time series, causing spectral contraction, over-smoothing, and poor mode coverage in time-series generation.

Approach

Augment the global FM velocity field with a bank of linear Koopman latent experts (dissipative parameterization) that receive gradients only through a WTA-selected expert per flow state, plus a load-balancing loss; at sampling, the selected expert's residual velocity is added to the global field.

Key results

• Context-FID on Sines/Stocks/ETTh/MuJoCo/Energy/fMRI: 0.003/0.027/0.015/0.006/0.022/0.155, best on 5 of 6 datasets vs. SMF and Diffusion-TS.
• Discriminative Score on Energy drops to 0.039 vs. 0.150 for second-best SMF; on fMRI 0.129 vs. 0.136.
• Ablation on Stocks: removing WTA loss raises Context-FID from 0.027 to 0.107; sampling with gamma=0 raises it to 0.091.

Surprising findings

• Predictive Score gains are minimal; several baselines match or beat PrismFlow on Predictive Score (e.g., SMF and Diffusion-TS on Stocks and fMRI).

Limitations

• Hard WTA routing may overuse few experts early in training, requiring a load-balancing regularizer.
• Performance gains are not uniform; Predictive Score and some Stocks and fMRI metrics favor baselines.
• Optimal number of experts depends on data complexity and no principled selection rule is given.

Experimental setup

Datasets: Sines Stocks ETTh MuJoCo Energy fMRI

Benchmarks: Sines Stocks ETTh MuJoCo Energy fMRI

Metrics: Context-FID Correlational Score Discriminative Score Predictive Score

Baselines: SMF Diffusion-TS DiffTime Diffwave TimeGAN TimeVAE Cot-GAN RNN-AR

Builds on: flow matching conditional flow matching Koopman operator theory Diffusion-TS SMF

Closest papers by meaning

• MissPath-FM: Flow Matching with Structured Priors for Partially Observed Time Series (NeurIPS 2026 · similarity 0.89)
• Elucidating Flow Matching ODE Dynamics via Data Geometry and Denoisers (ICML 2025 · similarity 0.88)
• SDFlow: Similarity-Driven Flow Matching for Time Series Generation (NeurIPS 2026 · similarity 0.87)
• Stream-level Flow Matching with Gaussian Processes (ICML 2025 · similarity 0.87)
• Diff2Flow: Training Flow Matching Models via Diffusion Model Alignment (CVPR 2025 · similarity 0.87)
• Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting (ICLR 2025 · similarity 0.87)