Tel Aviv researchers' new flow matching method cuts trajectory curvature by 2x

Flow matching models transport samples from a simple prior to a complex data distribution. When prior-data pairs are coupled via optimal transport (OT), trajectories become straight and non-crossing, enabling fast generation. However, computing OT couplings in high dimensions is intractable, and existing methods introduce bias or significant overhead. In a new paper, researchers from Tel Aviv University propose a clever reformulation: instead of solving for the OT coupling, treat the prior as a design choice. They identify low-frequency projection of natural images — essentially blurry versions of the data — as a prior that admits an OT-optimal identity coupling. This means the coupling between a blurry image and its full-resolution counterpart is already optimal, so the remaining flow matching task reduces to synthesizing high-frequency detail. Interpolating this prior with Gaussian noise further improves generation quality while preserving the OT coupling.

The approach requires no modifications to the flow model itself and integrates naturally with latent-space models, classifier-free guidance, and one-step generation frameworks. On all tested benchmarks, the method reduces trajectory curvature by more than 2x compared to existing flow matching techniques, yielding better generation quality in the few-step regime. The lightweight prior can be sampled with a small, fast model at inference time. For practitioners, this means faster image generation pipelines with higher fidelity, without the computational burden of optimal transport solvers. The work points to a deeper insight: the prior is not a fixed input but a powerful design lever that can drastically simplify generation.