One-Step Generation: The Frontier of Single-Step Generative Models

Overview

Diffusion Models and Flow Matching have achieved high-quality image generation, but they carry a fundamental limitation: inference requires tens to hundreds of iterative computation steps. This computational cost has made real-time applications and deployment on edge devices difficult.

From 2025 to 2026, methods that achieve high-quality generation in a single step (1-NFE) have been rapidly advancing. This series curates four papers driving this field in chronological order, tracing the technical evolution from extensions of Flow Matching to entirely new paradigms.

Common themes:

  • Eliminating dependence on distillation and pretrained models, enabling training from scratch
  • Formulations accompanied by theoretical guarantees (e.g., upper bounds on Wasserstein distance)
  • Quantitative evaluation via FID scores on ImageNet 256x256

Overview of Methods

flowchart TB
    FM["<b>Flow Matching (Baseline)</b><br/>Instantaneous velocity v(z_t, t)<br/>Multi-step ODE solver (50–250 NFE)<br/>FID ~2.27 (DiT-XL/2, 250 NFE)"]
    MF["<b>MeanFlow</b> (May 2025)<br/>Average velocity u<br/>1-NFE: FID 3.43"]
    TM["<b>Transition Matching</b> (Jun 2025)<br/>Discrete-time Markov transitions<br/>7x speedup"]
    TVM["<b>Terminal Velocity Matching</b> (Nov 2025)<br/>Terminal-time regularization<br/>1-NFE: FID 3.29 / 4-NFE: FID 1.99"]
    DM["<b>Drifting Models</b> (Feb 2026, Kaiming He)<br/>New paradigm: Evolve distribution during training<br/>1-NFE: FID 1.54"]

    FM --> MF
    FM --> TM
    FM --> TVM
    MF --> DM
    TVM --> DM
Figure 1: Lineage of one-step generation models. Starting from Flow Matching, three methods developed in parallel, and Drifting Models established a new paradigm.

Positioning of Each Paper

1. MeanFlow (May 2025)

While Flow Matching learns an “instantaneous velocity,” MeanFlow introduces a new quantity: the “average velocity.” The average velocity is displacement divided by a time interval, enabling direct sample generation in a single step.

Core idea: An identity called the MeanFlow Identity connects the relationship between the average velocity and the instantaneous velocity. This relationship enables training without explicitly computing integrals.

Results: Achieved FID 3.43 (1-NFE) on ImageNet 256x256. This was the best performance at the time for training from scratch without distillation or pretraining.

Details: MeanFlow

2. Transition Matching (June 2025)

Transition Matching is a framework that unifies diffusion models, Flow Matching, and autoregressive models as discrete-time Markov transitions. It proposes three variants (DTM, ARTM, FHTM), each exploring a different design space.

Core idea: The generative process is formulated as a sequence of stochastic transition kernels, with each transition matched independently. This enables flexible designs that differ from deterministic Flow Matching.

Results: DTM achieves a 7x speedup over Flow Matching (128 to 16 forward passes) while surpassing it in image quality and prompt alignment. FHTM is the first fully causal model to outperform Flow Matching.

Details: Transition Matching

3. Terminal Velocity Matching (November 2025)

This method generalizes Flow Matching by regularizing the velocity field at the terminal time of trajectories. While MeanFlow differentiates with respect to the start time, TVM differentiates with respect to the terminal time, obtaining stronger theoretical guarantees.

Core idea: By imposing a differential condition at the terminal time of the displacement map, an explicit upper bound on the 2-Wasserstein distance is derived. In practice, architectural modifications to ensure Lipschitz continuity (RMSNorm, QK-normalization) are key.

Results: Achieved FID 3.29 (1-NFE) and FID 1.99 (4-NFE) on ImageNet 256x256. At 4-NFE, it surpassed the performance of 500-NFE diffusion models.

Details: Terminal Velocity Matching

4. Generative Modeling via Drifting (February 2026)

An entirely new paradigm by Kaiming He et al. While conventional methods perform iterative “pushforward” at inference time, Drifting Models evolve the pushforward distribution during training. At inference time, only a single forward pass is required.

Core idea: A vector field called the “Drifting Field” attracts generated samples toward the data distribution while repelling them from other generated samples. Through anti-symmetry, the distribution naturally reaches equilibrium when it matches the target.

Results: Achieved FID 1.54 (latent space) / FID 1.61 (pixel space) on ImageNet 256x256. This established a new SOTA for 1-NFE generation.

Details: Drifting Models

Performance Comparison

Table 1: FID score comparison on ImageNet 256x256. Lower is better. Transition Matching is evaluated on a different dataset (Shutterstock 350M) and therefore not included in direct comparison.
Method Date 1-NFE FID 4-NFE FID No Distillation Key Feature
DiT (FM) 2023 - - - Baseline (250-NFE: 2.27)
MeanFlow 2025-05 3.43 - Yes Average velocity, from-scratch training
TVM 2025-11 3.29 1.99 Yes Terminal regularization, W2 upper bound
Drifting 2026-02 1.54 - Yes New paradigm, distribution evolution during training

Technical Background

Flow Matching Basics

Flow Matching learns a continuous transformation from a noise distribution \(p_0\) to a data distribution \(p_1\). A velocity field \(v(z_t, t)\) at time \(t \in [0, 1]\) is approximated by a neural network, and sampling is performed by solving an ODE:

\[ \frac{dz_t}{dt} = v_\theta(z_t, t) \]

At sampling time, starting from \(z_1 \sim p_1\) (noise), the ODE is solved in the reverse direction to obtain \(z_0\) (data). More steps yield higher accuracy, but at greater computational cost.

Challenges of One-Step Generation

One-step generation is equivalent to approximating the entire ODE trajectory with a single neural network evaluation. This inherently involves the following challenges:

  • Trajectory curvature: Non-linear trajectories are difficult to approximate in one step
  • Mode collapse: Maintaining diversity while preserving high quality is challenging
  • Consistency: Consistency across different scales must be guaranteed

Each paper in this series addresses these challenges with its own unique approach.

Future Prospects

The rapid evolution of one-step generative models suggests the following directions:

  • Real-time applications: Extension to video generation and interactive image editing
  • Robotics: Drifting Models have already demonstrated effectiveness as an alternative to Diffusion Policy
  • Multimodal integration: Transition Matching’s FHTM enables integration with LLM architectures
  • Theoretical understanding: A unified understanding of the theoretical relationships among these methods