Drift Flow Matching: Bridging Drift and Flow Models

• Drift Flow Matching (DFM) is a generative modeling paradigm that fuses one-step drift models with iterative flow matching to offer flexible sampling strategies.
• It utilizes a drift field parameterized transport map with kernel-weighted corrections, ensuring precise state transitions and bounded W2 geometry.
• Empirical results in image synthesis and control tasks demonstrate DFM’s effective trade-off between rapid generation and improved sample quality through multi-step refinement.

Drift Flow Matching (DFM) is a generative modeling paradigm that interpolates between the efficiency of one-step Drift Models and the test-time flexibility of iterative Flow Matching or Diffusion models. DFM enables both fast direct generation and refinement through multiple inference steps, combining unified theoretical underpinnings with strong empirical performance across domains such as image synthesis, latent generative modeling, and control. The core idea is to construct an adaptable transport map parameterized by a drift field, capable of functioning as either a single-step direct generator or a multi-step flow-based sampler (Ma et al., 17 May 2026).

1. Theoretical Framework

DFM generalizes and bridges two main classes of generative methods:

• Drift Models: Learn a deterministic mapping $f_\theta$ that directly transports samples from a base distribution $p_0$ to a target data distribution $p_1$, typically using a learned drift velocity field $V_{q,p}$ defined by kernel-weighted differences between $p_1$ and the model’s output $q$.
• Flow Matching: Learns a velocity field $v_\theta(x, t)$ such that solutions to the ODE $dx_t/dt = v_\theta(x_t, t)$, initialized from the base law $p_0$, match the law $p_1$ at $p_0$0. Training is via regression to analytically computable conditional velocities derived from pre-defined interpolants between endpoints.

Drift Flow Matching defines a two-time transport map

$p_0$1

where the neural vector field $p_0$2 learns to approximate the drift required to move any point $p_0$3 at time $p_0$4 to its location at a later time $p_0$5. DFM reduces to a one-step Drift Model when $p_0$6 and to Flow Matching in the infinitesimal-step limit as $p_0$7. At each pair of times $p_0$8, DFM constructs empirical marginals using pre-specified interpolation schedules ($p_0$9, $p_1$0), predicts the next state, and applies a kernelized drift correction to align predictions with the true target-time distribution (Ma et al., 17 May 2026).

2. Training Objectives and Algorithms

DFM training operates by grouping batch samples over pairs of times $p_1$1 and minimizing a stop-gradient loss reflecting the deviation between the group-wise predictions and drift-corrected targets: $p_1$2 Here, $p_1$3 is a group-wise kernel-weighted drift correction composed of positive (target) and negative (prediction) components, and $p_1$4 denotes a stop-gradient operator. Assigning each minibatch to groups under different $p_1$5 enables structured drift estimation that controls the dependence of the model path on the interpolation geometry.

Input representations explicitly embed absolute time and interval length ($p_1$6). The architecture is typically a residual MLP for low-dimensional or latent domains, with the same design extensible to high-dimensional settings. Kernel temperature and group size constitute key hyperparameters for stable long-step transport (Ma et al., 17 May 2026).

3. Inference Regimes and Quality–Efficiency Trade-off

At test time, DFM offers two key sampling regimes:

• One-step regime: Use $p_1$7 and apply $p_1$8 for rapid direct generation, matching the efficiency of classical Drift Models.
• Multi-step regime: March over a partition $p_1$9 and apply

$V_{q,p}$0

This recovers a forward-Euler integration of Flow Matching ODEs for small steps. Increasing the number of steps $V_{q,p}$1 enables monotonic improvement in quality, approaching or surpassing traditional Flow Matching methods.

Practical results show that even a small number of multi-step refinement steps allows DFM models to improve distributional coverage and sample fidelity well beyond the raw one-step output, delivering a tunable trade-off between computational efficiency and sample quality (Ma et al., 17 May 2026).

4. Geometric and Theoretical Properties

DFM inherits a controlled transport geometry from its interpolant choice: $V_{q,p}$2 Given linear schedules ($V_{q,p}$3, $V_{q,p}$4), the total discrete action along a partition is bounded by the endpoint second moment ($V_{q,p}$5). As $V_{q,p}$6, the DFM velocity field converges to that of standard Flow Matching (Ma et al., 17 May 2026).

Gradient flows induced by the DFM objective push the model distribution toward the target, and the method generalizes to new settings such as conditional modeling in latent spaces and control environments.

5. Empirical Performance and Applications

Drift Flow Matching demonstrates strong empirical results on diverse tasks:

Domain	1-step DFM performance	Multi-step DFM gain
2D synthetic datasets	Good initial coverage; smooth refinement with more steps	Multi-step path straightens flows, improves support
MNIST (latent VAE)	EMD ≈ 37.3, 100% acc. (1-step)	5 steps: EMD ≈ 37.2, 100% acc.
FFHQ (512d latent)	FID = 116.2 (1-step)	FID = 80.4 (5 steps), FID = 75.9 (10 steps)
ImageNet 256x256 (VAE)	FID = 1.52, IS = 259.4 (1-step)	FID = 1.31, IS = 287.4 (10 steps)
Robotic control	1-step nearly matches Diffusion Policy (100 steps) on simple tasks	5 steps yields further improvement on complex tasks

In all scenarios, DFM enables interpolation between efficiency and expressiveness, frequently achieving or exceeding the quality of true Flow Matching or Diffusion models at much lower inference cost (Ma et al., 17 May 2026).

6. Technical Limitations and Ongoing Research

DFM requires estimating drift corrections via kernel-based groupings, leading to $V_{q,p}$7 computational cost per group, which can become restrictive for very large minibatches. The method’s expressiveness depends on kernel temperature and sufficient group size; insufficient group size degrades performance in large-step regimes. In finite-sample domains, drift identifiability remains heuristic.

Notable theoretical properties include the inheritance of $V_{q,p}$8-geometry from the flow matching path and guarantees that gradient descent on the DFM loss aligns the model distribution with the target under mild regularity. Theoretical and empirical ablations reveal robustness to architectural variations and highlight the importance of appropriate time-embedding and group sampling strategies.

Promising extensions include the use of alternative drift definitions (e.g., Sinkhorn or feature-space drifts), adaptive time-pair samplers, and applications to conditional and high-dimensional generative models. Further theoretical analysis of convergence and stability under finite-sample drift estimation is an open direction (Ma et al., 17 May 2026).

7. Relationship to Broader Flow Matching Literature

DFM generalizes the core mechanism of Flow Matching to enable continuous interpolation between Drift Models and classical auto-regressive, ODE- or SDE-driven flow matching approaches. It retains key theoretical benefits, such as $V_{q,p}$9-action bounds and monotone improvement with increased inference steps, but uniquely accommodates efficiency-quality trade-offs through its tunable inference regime. The method has been concretely instantiated across latent generative modeling, robotic control, and high-dimensional data domains, demonstrating both strong performance and adaptability.

Code and pre-trained checkpoints are announced for public release, and full mathematical proofs are made available alongside training and inference pseudocode in the main manuscript (Ma et al., 17 May 2026).