![[2602.04908] Temporal Pair Consistency for Variance-Reduced Flow Matching](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.04908] Temporal Pair Consistency for Variance-Reduced Flow Matching
Summary
The paper introduces Temporal Pair Consistency (TPC), a novel approach to reduce variance in flow matching for continuous-time generative models, enhancing sample quality and efficiency without altering existing architectures.
Why It Matters
This research addresses the challenge of high estimator variance in generative models, which can hinder performance and efficiency. By proposing TPC, the authors provide a method that improves model training outcomes, making it relevant for advancements in machine learning and computer vision.
Key Takeaways
- Temporal Pair Consistency (TPC) couples velocity predictions at paired timesteps to reduce variance.
- TPC operates at the estimator level, maintaining existing model architectures and solvers.
- The method shows improved sample quality and efficiency on datasets like CIFAR-10 and ImageNet.
- TPC achieves lower Fréchet Inception Distance (FID) at comparable or reduced computational costs.
- The approach integrates seamlessly with state-of-the-art training pipelines.
Computer Science > Machine Learning arXiv:2602.04908 (cs) [Submitted on 4 Feb 2026 (v1), last revised 19 Feb 2026 (this version, v2)] Title:Temporal Pair Consistency for Variance-Reduced Flow Matching Authors:Chika Maduabuchi, Jindong Wang View a PDF of the paper titled Temporal Pair Consistency for Variance-Reduced Flow Matching, by Chika Maduabuchi and 1 other authors View PDF HTML (experimental) Abstract:Continuous-time generative models, such as diffusion models, flow matching, and rectified flow, learn time-dependent vector fields but are typically trained with objectives that treat timesteps independently, leading to high estimator variance and inefficient sampling. Prior approaches mitigate this via explicit smoothness penalties, trajectory regularization, or modified probability paths and solvers. We introduce Temporal Pair Consistency (TPC), a lightweight variance-reduction principle that couples velocity predictions at paired timesteps along the same probability path, operating entirely at the estimator level without modifying the model architecture, probability path, or solver. We provide a theoretical analysis showing that TPC induces a quadratic, trajectory-coupled regularization that provably reduces gradient variance while preserving the underlying flow-matching objective. Instantiated within flow matching, TPC improves sample quality and efficiency across CIFAR-10 and ImageNet at multiple resolutions, achieving lower FID at identical or lower computational ...