PaperlayerView exact revision on arXiv

Source-linked AI summary

Temporal Pair Consistency for Variance-Reduced Flow Matching

Chika Maduabuchi, Jindong Wang

arXiv:2602.04908v2cs.LGcs.AIcs.CV

TL;DR

Continuous-time generative models learn temporal dynamics independently, leaving correlated evaluations with high variance and reduced sampling efficiency. TPC couples velocity predictions at paired timesteps along the same probability path, improving sample quality and efficiency across flow-matching and rectified-flow settings without changing paths, solvers, or architectures.

  • Problem

    Standard flow matching learns velocity predictions independently across time, despite shared randomness along probability paths, limiting temporal coherence and sample efficiency.

  • Method

    TPC pairs stochastic evaluations at two timesteps along the same probability path, enforcing velocity consistency as a variance-reducing control-variate estimator.

  • Results

    TPC consistently improves sample quality and sampling efficiency across flow matching, rectified flow, and modern noise-augmented, score-based pipelines.

  • Takeaways & Limitations

    Simple temporal coupling can replace more complex path or solver designs within the evaluated continuous-time image-generation settings.

  • Takeaways & Limitations

    The study focuses on unconditional image generation at resolutions up to 128 × 128; conditional settings, higher resolutions, and other modalities remain future work.

Abstract

from arXiv · show

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 cost than prior methods, and extends seamlessly to modern SOTA-style pipelines with noise-augmented training, score-based denoising, and rectified flow.

1. Introduction

Temporal Pair Consistency (TPC) reduces variance in continuous-time generative-model training by coupling velocity evaluations at paired timesteps along the same probability path. It preserves the standard flow-matching objective while improving sample quality and sampling efficiency across flow matching and rectified flow.

  • Motivation: Standard flow matching learns velocity predictions independently across time, increasing gradient variance and numerical error under coarse inference discretizations.These effects reduce sample efficiency by requiring finer discretization or more function evaluations, even with identical solvers and probability paths.
  • Method: TPC couples stochastic velocity evaluations at paired timesteps along the same probability path to enforce temporal coherence and reduce estimator variance.The method operates within the standard flow-matching objective without modifying the model architecture, probability path, or solver.
  • Method: Fixed antithetic and learnable monotone timestep-pairing mechanisms provide practical ways to construct temporal correspondences.The learnable mechanism adapts coupling to the data and model while preserving the ordered structure of the probability path.
  • Theory: TPC induces a quadratic, trajectory-coupled regularizer while preserving the underlying flow-matching objective.Temporal consistency is applied stochastically, making TPC a variance-reduction mechanism rather than a hard constraint.
  • Evaluation: TPC improves sample quality and sampling efficiency across flow matching and rectified flow on CIFAR-10 and ImageNet benchmarks.The gains extend to standard probability-flow sampling and modern pipelines using noise-augmented training and score-based denoising.

2. Related Work

TPC differs from prior temporal, trajectory, and consistency regularization by coupling paired stochastic flow-matching evaluations rather than explicitly constraining dynamics or enforcing output invariance. It is complementary to methods that design probability paths and trajectory geometry.

  • Temporal and trajectory regularization: Prior continuous-time methods regularize learned ODE dynamics through path length, kinetic energy, or velocity-field Jacobians, often targeting stability or invertibility.These approaches include explicit temporal smoothness, consistency, and trajectory regularization.
  • Temporal and trajectory regularization: TPC instead couples velocity predictions at paired timesteps through stochastic evaluations of the existing flow-matching objective, without explicit time-smoothness, Jacobian, or higher-order-derivative penalties.This distinguishes TPC from path- and Jacobian-based regularization methods.
  • Consistency regularization: Unlike consistency regularization and consistency models, TPC neither enforces output invariance nor defines a fixed-point mapping across time.It reduces training estimator variance while preserving the underlying continuous-time dynamics.
  • Relation to flow matching and rectified flow: TPC is complementary to flow matching, rectified flow, and other probability-path constructions because it targets temporal estimator structure rather than path choice or trajectory geometry.The cited methods focus on designing efficient probability paths and shaping trajectories.

3. Method

TPC-FM reduces flow-matching estimator variance by coupling velocity evaluations at paired timesteps along a shared probability path. It preserves the underlying probability path, solver, and flow-matching formulation while inducing temporal regularization through the estimator.

  • Temporal Pair Consistency: TPC couples stochastic velocity evaluations at two timesteps sharing endpoint randomness, converting correlated gradient noise into a variance-reduction mechanism.The paired estimator increases gradient correlation and produces a control-variate effect.
  • Pairing Operators: Fixed antithetic pairing uses ψ_fix(t) = 1 − t, causing paired gradients to tend toward negative correlation and reducing estimator variance without additional parameters.This construction exploits time-reversal symmetry for commonly used symmetric interpolants.
  • Pairing Operators: Learned monotone pairing uses a low-capacity map ϕ with ϕ′(t) ≥ 0, while a K = 32 grid penalty weakly discourages order violations.The learned map enables adaptive pairings while preserving temporal order.
  • Temporal Pair Consistency: The temporal consistency penalty is a quadratic term ∥v_t − v_t′∥2 that enforces coherence between paired velocity predictions along the same path.At the population level, this acts as a Tikhonov-regularized risk selecting fields with reduced temporal oscillation.
  • Estimator Design: TPC operates entirely at the estimator level and preserves the conditional flow-matching formulation, with stochastic gating limiting over-regularization and maintaining unregularized gradient exposure.The gating variable b follows Bernoulli(p_tpc), yielding E[˜ℓ_TPC] = p_tpcℓ_TPC.
  • Theoretical Analysis: Under mild regularity, optimal coupling gives Var(g − α⋆g′) = Var(g)(1 − ρ2), certifying strict gradient-variance reduction when paired gradients are positively correlated.The analysis connects this reduction to improved optimization at fixed step size or reduced NFE for a target error.

4. Experiments

Experiments evaluate TPC across CIFAR-10 and ImageNet at multiple resolutions, including flow matching, rectified flow, and modern noise-augmented SOTA-style protocols. TPC consistently improves sample quality and efficiency, while remaining robust to hyperparameter choices.

  • Benchmark setup: Experiments cover unconditional CIFAR-10 and ImageNet generation at 32×32, 64×64, and 128×128 resolutions under established diffusion and flow-modeling protocols.Comparisons include state-of-the-art diffusion- and flow-based models.
  • SOTA-style evaluation: Modern ImageNet evaluations use noise-augmented training with score-based denoising at sampling time, matching protocols used by competitive high-resolution generative pipelines.Baseline FM and TPC-FM use identical noise configurations unless otherwise stated, enabling controlled comparisons.
  • Flow Matching: TPC-FM reduces CIFAR-10 FID from 6.35 for FM w/ OT to 3.19 at the same NFE, while matching the best reported flow-based NLL.Across datasets and resolutions, TPC-FM improves prior flow-matching objectives at identical or lower NFE with competitive likelihoods.
  • Rectified Flow: TPC-RF consistently improves FID and recall at identical NFE across rectification depths in one-step and full-simulation rectified-flow evaluations.The comparison includes rectified flow, probability-flow ODE, and SDE baselines.
  • Hyperparameter analysis: Across CIFAR-10 hyperparameter settings, TPC-FM improves over the FM w/ OT baseline with FID 6.35, while moderate coupling performs best and excessive λtpc degrades quality.The analysis varies ptpc, λtpc, and λmono, showing robustness alongside a cost to overly strong temporal constraints.

5. Conclusion

TPC is presented as a lightweight, theoretically grounded variance-reduction principle that couples velocity predictions across paired timesteps. It improves sample quality and efficiency while remaining applicable to modern training pipelines, though evaluation is limited to unconditional image generation up to 128 × 128.

  • TPC enforces coherence between velocity predictions at paired timesteps along the same probability path, with theoretical variance-reduction guarantees.The principle is applied to flow matching and rectified flow.
  • TPC consistently improves sample quality and efficiency in both one-step and full-simulation regimes.Its effectiveness extends to modern SOTA-style training pipelines.
  • TPC’s simple temporal coupling can replace more complex probability-path or solver designs.
  • The study covers unconditional image generation at resolutions up to 128 × 128.Conditional settings, higher resolutions, and other modalities remain future work.

A. Theoretical Analysis · A.1. Preliminaries · A.2. TPC-FM as quadratic regularization in a trajectory-coupled Hilbert space

The analysis frames flow matching as independent-in-time L2 regression, then shows that TPC adds trajectory-coupled quadratic regularization selecting temporally coherent predictors without changing the underlying population objective.

  • A.1. Preliminaries: An admissible probability path connects p0 and p1 through a measurable velocity field satisfying the continuity equation.
  • A.1. Preliminaries: The path sampler jointly samples endpoint pairs and time, then maps them to states through Φ to define the trajectory geometry.
  • A.1. Preliminaries: Flow matching defines a population risk for measurable predictors against conditional target velocities.
  • A.1. Preliminaries: Lemma A.1 identifies the conditional target velocity as the unique L2 projection minimizer under G_t-measurable predictors.
  • A.1. Preliminaries: Because the baseline objective fits each timestep independently, minimal population risk does not prevent arbitrary temporal oscillations in the learned vector field.
  • A.2. TPC-FM as quadratic regularization in a trajectory-coupled Hilbert space: TPC regularizes prediction discrepancies at paired states generated from the same sampled trajectory, rather than imposing abstract smoothness in time.
  • A.2. TPC-FM as quadratic regularization in a trajectory-coupled Hilbert space: The TPC seminorm is the L2(π ⊗ ρ) norm of the pairing operator, providing a trajectory-coupled Hilbert-space geometry.
  • A.2. TPC-FM as quadratic regularization in a trajectory-coupled Hilbert space: Lemma A.2 shows that regularization selects near-minimal-risk predictors with smaller temporal variation along coupled trajectories.

A.3. Anchor theorem: population regularization ⇒correlated gradients ⇒strict variance reduction

The anchor theorem unifies population regularization, positive correlation between gradients at paired timesteps, and strict variance reduction under temporal coupling. It derives this coupling from smooth time dependence and shared path randomness, then relates temporal coherence to probability-flow sampling behavior.

  • Gradient correlation: TPC correlates gradient evaluations by pairing timesteps while sharing the randomness that generates the path and sampler noise.The paired time is defined as t′ = ψ(t), with h(t, ξ) := g(t′, ξ).
  • Variance reduction: Positive covariance enables an optimal scalar control variate to reduce the variance of G − αH strictly below the unadjusted gradient variance.The sufficient correlation condition is derived from smoothness and path coupling, with the strict reduction using the lower bound on covariance.
  • Anchor theorem: Theorem A.5 consolidates population contraction, quantitative gradient correlation, and strict variance reduction as consequences of one temporal coupling mechanism.The theorem assumes smoothness conditions and a pairing with |t − ψ(t)| ≤ ∆ almost surely.
  • Assumptions: The positive-correlation requirement follows from smooth dependence on time and path coupling induced by shared endpoints rather than an ad-hoc assumption.The theorem explicitly links this condition to the pairing distance bound |t − ψ(t)| ≤ ∆.
  • Sampling implications: Trajectory-wise temporal coherence is the relevant notion for numerical behavior because probability-flow sampling integrates the learned ODE only along sampler-visited states.This connects the theorem’s temporal regularity to generation-time numerical behavior.

A.4. ODE link: TPC as control of temporal roughness and solver error

TPC links temporal smoothness of the learned vector field to numerical integration: under Lipschitz and time-differentiability assumptions, temporal variation explicitly contributes to Euler discretization error. Local pairings control trajectory-averaged temporal derivatives, while global pairings suppress low-frequency temporal oscillations on sampler trajectories.

  • Global error bound: Under Lipschitz-in-x and time-differentiability assumptions, Lemma A.6 makes temporal variation of the vector field an explicit term in the global Euler error.The proof uses a time Taylor expansion, Lipschitz bounds, and Grönwall accumulation.
  • Solver stability: Suppressing temporal roughness along sampled trajectories improves numerical stability at fixed solver step size.This follows because temporal variation is isolated as a direct contributor to global discretization error.
  • Local pairings: For local pairings, small TPC discrepancy yields a small trajectory-averaged temporal derivative, the quantity entering time-variation terms in discretization bounds.The local expansion relates paired velocity differences to temporal and state-gradient variation along the trajectory.
  • Global pairings: For global pairings such as antithetic t ↦ 1 − t, TPC controls low-frequency temporal oscillations spectrally and reduces effective roughness on sampler support.The control is expressed through the quadratic form induced by Aψ.

A.5. Uniform convergence under TPC constraint

Under the TPC constraint, the predictor class is restricted by a norm threshold, enabling a Rademacher-complexity generalization bound. In typical parameterizations, the constrained class’s complexity does not increase with the threshold parameter τ.

  • Constraint definition: The constrained class V_τ consists of predictors v ∈ V satisfying the stated TPC norm constraint, with predictors bounded by ∥v(x, t)∥₂ ≤ B almost surely.The passage defines V_τ through a norm-based constraint on predictors from the bounded class V.
  • Uniform convergence bound: Proposition A.7 provides a Rademacher-complexity bound for the constrained class V_τ with probability at least 1 − δ.The bound is stated for universal constants and vector-valued Rademacher complexity.
  • Complexity behavior: In typical parameterizations, R_n(V_τ) is non-increasing in τ.Increasing the constraint threshold does not increase the constrained class’s Rademacher complexity under these parameterizations.

A.6. Limitations · A.7. Takeaway

The analysis certifies mechanism-level reductions rather than monotone FID/IS improvement, while showing that temporal correlation and local TPC penalties reduce gradient variance, temporal roughness, and trajectory-averaged discretization error under stated assumptions.

  • A.6. Limitations: TPC’s guarantees do not imply monotone improvement in FID or IS.The reported results certify mechanism-level reductions only, not monotone metric improvement.
  • A.6. Limitations: The analysis identifies model bias, optimization error, solver error, and estimation noise as limitations.
  • A.6. Limitations: Higher correlation between paired velocity predictions implies lower SGD-gradient variance and lower temporal roughness on sampler support, which reduces ODE discretization error.
  • A.7. Takeaway: Under the stated assumptions, paired-time gradients are represented as G = g(t, ξ) and H = g(ψ(t), ξ).
  • A.7. Takeaway: The correlation parameter ρ is quantitatively lower bounded via Lemma A.4 under path-coupled Lipschitz regularity.
  • A.7. Takeaway: For ˙zt = vθ(zt, t), Euler-type global error explicitly depends on temporal variation, which local TPC penalties control in a trajectory-averaged sense through (20).

B. Generated Samples

The section compares GAN, ODE, and SDE generative models on CIFAR-10 under one-step and full-simulation settings, while showing that TPC-FM rapidly collapses training variance and remains stable.

  • Model comparisons: Table 8 compares GAN, ODE, and SDE generative models on CIFAR-10.The comparison covers both one-step and full-simulation settings.
  • Model comparisons: Table 8 evaluates the models under one-step generation settings.The table explicitly includes one-step simulation as a comparison setting.
  • Training variance: TPC-FM shows an early variance collapse and sustained stability throughout training.Figure 4 reports the training-variance behavior of TPC-FM.
Loading 2602.04908v2…