PrismFlow: New Time-Series Generator Beats State-of-the-Art by 38.6%

Generating realistic time-series data is notoriously difficult due to multimodal patterns and multiscale dynamics like oscillations and high-frequency variations. Standard Flow Matching (FM) methods rely on a single global vector-field estimator, which can lead to overly smoothed approximations when local dynamics require incompatible conditional velocities. This smoothing attenuates branch-specific dynamics, causing spectral distortion and poor mode coverage.

To overcome this, a team led by Junru Zhang proposes PrismFlow, a novel FM approach that introduces Koopman-inspired dynamical experts. Each expert learns residual corrections in a latent space where local nonlinear temporal evolution can be approximated by linear transitions. A confidence-aware Winner-Take-All (WTA) objective updates only the expert best aligned with each sample, encouraging mode-specific specialization. During sampling, the selected expert adds a residual dynamical correction to the global transport field, preserving FM stability while recovering fine-grained temporal structures.

Across multiple benchmarks, PrismFlow delivers significant improvements: a 15.6% gain in Context-FID and a 38.6% improvement in Discriminative Score over standard FM. The method also performs robustly in low-data settings and is effective for forecasting and imputation tasks. This work bridges physics-inspired Koopman theory with modern generative modeling, offering a practical solution for high-fidelity time-series generation.