Skip to content
BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
.gitignore
.gitignore
 
 
CONTRIBUTING.md
CONTRIBUTING.md
 
 
LICENSE
LICENSE
 
 
README.md
README.md
 
 

Repository files navigation

Awesome One-Step Generation

Awesome License: MIT PRs Welcome

"Nature is pleased with simplicity."
— Isaac Newton

A curated list of papers, code, and resources for one-step diffusion models — methods that turn noise into high-quality samples in a single neural network forward pass.

🧭 Classification Rationale

Instead of the usual native vs. distillation split, we group methods by the core mathematical principle that collapses iterative sampling into one step. Each paper is placed by its primary contribution, with cross-references for works that span multiple principles.

Legend Meaning
🟢 Native Trained from scratch / teacher-free; one-step is a native training objective
🔵 Distill Distills a pretrained multi-step diffusion teacher into one/few steps
🟡 Hybrid Supports both native training and distillation

Table of Contents

Consistency Models

Core principle — self-consistency along the PF-ODE. Any two points on the same probability-flow ODE trajectory must map to the same endpoint. Enforcing this consistency constraint lets a network jump directly from noise to data. Following iCT, the principle admits both consistency training (CT, 🟢 native) and consistency distillation (CD, 🔵 teacher-based).

Native (🟢)

  • Consistency Models [ICML 2023] 🟡
    [Paper]
    Self-consistency for direct noise-to-data mapping; supports both distillation (CD) and native training (CT).

  • Improved Techniques for Training Consistency Models (iCT) [ICLR 2024] 🟢
    [Paper]
    Makes consistency training from scratch competitive with distillation.

  • Consistency Models Made Easy [ICLR 2025] 🟢
    [Paper]
    Simplified and efficient consistency-model training recipe.

  • Inverse Flow and Consistency Models [ICML 2025] 🟢
    [Paper]
    Inverse-flow approach for consistency model training.

  • How to Build a Consistency Model: Learning Flow Maps via Self-Distillation [NeurIPS 2025] 🟡
    [Paper] [Code]
    A systematic approach to learning flow maps via self-distillation.

Distillation (🔵)

  • Latent Consistency Models (LCM) [ICLR 2024] 🔵
    [Paper]
    Consistency distillation in latent space from pretrained Stable Diffusion.

  • Consistency Trajectory Models (CTM) [ICLR 2024] 🔵
    [Paper] [Code]
    Learning the probability-flow ODE trajectory for consistency.

  • SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation [ICCV 2025] 🔵
    [Paper]
    Continuous-time consistency distillation for ultra-fast text-to-image.

  • Simple Distillation for One-Step Diffusion Models (CED) [NeurIPS 2025] 🔵
    [Paper]
    Contrastive energy distillation for one-step diffusion.

  • Phased Consistency Models (PCM) [NeurIPS 2024] 🔵
    [Paper] [Code]
    Phased consistency distillation that surpasses LCM; supports 1-16 step generation.

  • Hyper-SD: Trajectory Segmented Consistency Model [NeurIPS 2024] 🔵
    [Paper] [Code]
    Trajectory-segmented consistency distillation with human feedback; surpasses SDXL-Lightning.

  • Truncated Consistency Models [2024] 🟢
    [Paper]
    Operating on truncated input time ranges for improved one-step and two-step quality.

  • Any-step Generation via N-th Order Recursive Consistent Velocity Field (RCGM) [ICLR 2026] 🟢
    [Paper] [Code]
    N-th order recursive consistent velocity field estimation supporting any-step generation.

Flow Straightening

Core principle — straighten the transport path. If the trajectory connecting noise and data is (close to) a straight line, a single Euler step approximates the full ODE integration. Rectified-flow and flow-matching methods learn or refine such straight paths so that one-step Euler ≈ the complete ODE solution.

Native (🟢)

Rectified Flow and Flow Matching are catalogued under Related Resources → Foundational Works as general frameworks. The methods below specifically target or achieve one-step generation.

  • Optimal Flow Matching: Learning Straight Trajectories in Just One Step [NeurIPS 2024] 🟢
    [Paper] [Code]
    Restricting flow matching to obtain one-step straight trajectories.

  • Improving the Training of Rectified Flows [NeurIPS 2024] 🟢
    [Paper] [Code]
    Improved rectified-flow training enabling stronger one-step performance.

  • End-to-End Single-Step Flow Matching via Direct Models (FlowFit) [ICLR 2025] 🟢
    [Paper]
    Direct model learning for single-step flow matching.

  • Mean Flows for One-step Generative Modeling (MeanFlow) [NeurIPS 2025] 🟢
    [Paper] [Code]
    Native one-step generation via mean-velocity-field properties; trains from scratch.

  • One-step Latent-free Image Generation with Pixel Mean Flows (pMF) [2026] 🟢
    [Paper]
    Latent-free one-step generation in pixel space; decouples output space from loss space, reaching 2.22 FID on ImageNet 256×256 and 2.48 FID on 512×512.

  • SoFlow: Solution Flow Models for One-Step Generative Modeling [ICLR 2026] 🟢
    [Paper] [Code]
    Learning the solution map of flow ODEs for direct one-step generation.

  • Rectified Diffusion: Straightness Is Not Your Need in Rectified Flow [ICLR 2025] 🟢
    [Paper] [Code]
    Challenges the straightness assumption; proposes a broader rectified flow framework.

  • Variational Flow Matching (S-VFM) [CVPR 2026] 🟢
    [Paper]
    Learning straight flows via variational latent codes integrated with the flow-matching objective.

  • Beyond Optimal Transport: Model-Aligned Coupling for Flow Matching [CVPR 2026 Findings] 🟢
    [Paper] [Code]
    Model-aligned coupling selects learnable source-target pairs to improve one-step/few-step flow matching generation.

  • SubFlow: Sub-mode Conditioned Flow Matching for Diverse One-Step Generation [arXiv 2026] 🟢
    [Paper]
    Sub-mode conditioning reduces averaging distortion and improves diversity in one-step flow matching.

Distillation (🔵)

Progressive Distillation is catalogued under Related Resources → Foundational Works as it is a general step-compression method rather than a one-step-specific technique. It is cross-referenced here as it inspired much of the flow-distillation lineage.

  • Bi-Anchor Interpolation Solver for Accelerating Generative Modeling [ICML 2026] 🔵
    [Paper] [Code]
    Lightweight training-based solver for flow matching models.

  • InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation [ICLR 2024] 🔵
    [Paper] [Code]
    Rectified-Flow reflow technique for one-step text-to-image.

  • Training Smaller One-Step Diffusion Models with Rectified Flow (SlimFlow) [ECCV 2024] 🔵
    [Paper] [Code]
    Parameter-efficient one-step models via rectified-flow compression.

  • Learning Few-Step Diffusion Models by Trajectory Distribution Matching (TDM) [ICCV 2025] 🔵
    [Paper] [Code]
    Unified trajectory and distribution matching for few-step distillation.

  • Align Your Flow: Scaling Continuous-Time Flow Map Distillation [NeurIPS 2025] 🔵
    [Paper]
    Continuous-time objectives for training flow maps at scale.

  • TRACT: Denoising Diffusion Models with Transitive Closure Time-Distillation [2023] 🔵
    [Paper] [Code]
    Transitive closure time-distillation for progressive step reduction.

  • Piecewise Rectified Flow (PeRFlow) [NeurIPS 2024] 🔵
    [Paper] [Code]
    Universal plug-and-play accelerator via piecewise linearization of rectified flows.

Distribution Divergence Minimization

Core principle — minimize a statistical divergence between distributions. These methods train the one-step generator so that its output distribution p_θ matches the target distribution p_target by minimizing some divergence D(p_θ ∥ p_target). As shown by Uni-Instruct (NeurIPS 2025), DMD, SiD, Diff-Instruct and many others are special cases of a single diffusion divergence minimization framework; they differ chiefly in which divergence they minimize. We organize them along that axis.

Forward KL / Regression

Minimizing the forward KL D(p_target ∥ p_θ) is mode-covering and reduces to maximum-likelihood / regression-style objectives.

  • Progressive Distillation is catalogued under Related Resources → Foundational Works, since its core contribution is step compression rather than a divergence-minimization result. It is cross-referenced here because its regression-on-teacher objective is forward-KL in spirit.

  • BOOT: Data-free Distillation of Denoising Diffusion Models with Bootstrapping [ICML 2023] 🔵
    [Paper]
    Data-free bootstrapping distillation via sequential prediction on the signal ODE.

  • EM Distillation for One-step Diffusion Models [NeurIPS 2024] 🔵
    [Paper]
    Maximum-likelihood (forward-KL) distillation via an Expectation-Maximization formulation.

Reverse KL / Distribution Matching

Minimizing the reverse KL D(p_θ ∥ p_target) is mode-seeking; in practice it is estimated through the score functions of the real and the generator ("fake") distributions.

  • One-step Diffusion with Distribution Matching Distillation (DMD) [CVPR 2024] 🔵
    [Paper]
    Learning score functions of real and fake distributions for reverse-KL distribution matching.

  • Improved Distribution Matching Distillation (DMD2) [NeurIPS 2024 Oral] 🔵
    [Paper] [Code]
    Improved DMD that removes the costly regression loss.

  • One-step Diffusion Models with f-Divergence Distribution Matching (f-distill) [2025] 🔵
    [Paper]
    A generalized f-divergence framework subsuming reverse-KL distribution matching.

Fisher Divergence / Score Matching

Matching the scores (gradients of log-density) of the generator and target corresponds to minimizing a Fisher-type divergence — the basis of score-identity and score-implicit methods.

  • Score identity Distillation (SiD) [ICML 2024] 🔵
    [Paper] [Code]
    Data-free score-identity matching for exponentially fast distillation.

  • One-Step Diffusion Distillation through Score Implicit Matching (SIM) [NeurIPS 2024] 🔵
    [Paper] [Code]
    Score implicit matching that preserves the teacher's denoising capability.

  • DiffRatio: Training One-Step Diffusion Models Without Teacher Supervision [ICLR 2025] 🟢
    [Paper] [Code]
    Direct score-ratio estimation for teacher-free one-step training.

  • Adversarial Score identity Distillation (SiDA) [ICLR 2025] 🔵
    [Paper] [Code]
    Adversarial enhancement of SiD, surpassing the teacher model.

  • Score-of-Mixture Training [ICML 2025 Spotlight] 🟢
    [Paper]
    Minimizing mixture-distribution divergence for pretrain-free one-step generation.

Variational Bounds

When the divergence cannot be optimized directly, a tractable variational bound (e.g. variational score distillation) is minimized instead, using a learned score/critic to guide the generator.

  • Diff-Instruct: A Universal Approach for Transferring Knowledge from Pre-trained Diffusion Models [NeurIPS 2023] 🔵
    [Paper]
    Universal framework using variational bounds to guide any generator; conceptual precursor to many one-step methods.

  • SwiftBrush: One-Step Text-to-Image Diffusion Model with Variational Score Distillation [CVPR 2024] 🔵
    [Paper] [Code]
    Image-free variational score distillation for one-step text-to-image.

  • SwiftBrush v2: Make Your One-step Model Better Than Its Teacher [ECCV 2024] 🔵
    [Paper] [Code]
    Improved SwiftBrush that surpasses teacher-model performance.

  • Diff-Instruct++: Training One-step Text-to-image Generator Model to Human-Preferred [2024] 🔵
    [Paper] [Code]
    Human-preference alignment for one-step text-to-image generators.

  • WaDi: Weight Direction-aware Distillation for One-step Image Synthesis [CVPR 2026] 🔵
    [Paper] [Code]
    Weight-direction-aware VSD framework with LoRaD adapters, improving one-step distillation with about 10% trainable parameters.

Adversarial Training

Core principle — push samples onto the data manifold with a discriminator. Borrowing the GAN principle, an adversarial discriminator drives generated samples toward the real-data distribution, yielding sharp, high-fidelity outputs in a single step. Often combined with score- or consistency-based objectives.

  • You Only Sample Once (YOSO) [ICLR 2024] 🟢
    [Paper] [Code]
    Self-cooperative diffusion-GAN for one-step text-to-image, trained without a frozen teacher.

  • Adversarial Diffusion Distillation (ADD) [ECCV 2024] 🔵
    [Paper]
    Adversarial training for distilling diffusion into 1–4 step generators (SDXL-Turbo).

  • Latent Adversarial Diffusion Distillation (LADD) [ECCV 2024] 🔵
    [Paper]
    Latent-space adversarial distillation overcoming ADD limitations (powers FLUX.1 [schnell]).

  • SDXL-Lightning: Progressive Adversarial Diffusion Distillation [2024] 🔵
    [Paper] [Code]
    Combines progressive and adversarial distillation for one-step 1024px text-to-image on SDXL.

  • TwinFlow: Realizing One-step Generation with Self-adversarial Flows [ICLR 2026] 🟢
    [Paper] [Code]
    Self-adversarial flow framework without fixed auxiliary models for one-step generation.

Shortcut / Direct Mapping

Core principle — learn the noise-to-data map directly. Rather than reasoning about trajectories or iterative denoising, these methods propose a new paradigm: a single conditioned network that maps noise straight to data, bypassing the ODE/SDE machinery entirely.

  • One Step Diffusion via Shortcut Models [ICLR 2025 Oral] 🟢
    [Paper] [Code]
    A single network and single training phase for direct generation without iterative denoising.

  • High-Order Matching for One-Step Shortcut Diffusion Models [ICLR 2025] 🟢
    [Paper]
    High-order matching to improve shortcut models.

  • Revisiting Diffusion Models: From Generative Pre-training to One-Step Generation (D2O) [ICML 2025] 🟢
    [Paper] [Code]
    Treating diffusion as generative pre-training and fine-tuning for one-step generation.

  • Noise Consistency Training (NCT) [NeurIPS 2025] 🟢
    [Paper] [Code]
    Native integration of new control signals directly into one-step generators.

  • Di[M]O: Distilling Masked Diffusion Models into One-step Generator [ICCV 2025] 🔵
    [Paper] [Code]
    First work distilling discrete masked diffusion models into a one-step generator.

Drifting Models

Core principle — evolve the generator's output distribution during training via Wasserstein gradient flow. Instead of iterating at inference time, drifting models introduce a "drifting field" that continuously pushes the generator's output distribution toward the data distribution during training. At inference, a single forward pass suffices. Theoretically equivalent to minimizing KL divergence via Wasserstein-2 gradient flow and secretly performing score matching under Gaussian kernels.

  • Generative Modeling via Drifting [2026] 🟢
    [Paper] [Code]
    The foundational drifting model: evolves generator distribution at training time; achieves FID 1.54 on ImageNet 256×256 with one-step generation.

  • Generative Drifting is Secretly Score Matching: a Spectral and Variational Perspective [2026] 🟢
    [Paper]
    Proves drifting under Gaussian kernels equals score difference of smoothed distributions; connects to JKO discretization.

  • Gradient Flow Drifting: Generative Modeling via Wasserstein Gradient Flows [2026] 🟢
    [Paper]
    Establishes exact equivalence between drifting and Wasserstein gradient flow of KL divergence; extends to Riemannian manifolds.

  • W-Flow: One-Step Generative Modeling via Wasserstein Gradient Flows [2026] 🟢
    [Paper] [Code]
    Trains a one-step generator with training dynamics guided by a Wasserstein gradient flow of the Sinkhorn divergence; achieves FID 1.29 on ImageNet 256×256 with one-step generation.

Applications

Text-to-Image / Personalization

  • PixelRush: Ultra-Fast, Training-Free High-Resolution Image Generation via One-step Diffusion [CVPR 2026] 🔵
    [Paper]
    Training-free patch-based low-step/one-step high-resolution text-to-image generation, targeting practical 4K synthesis.

  • Adversarial Concept Distillation for One-Step Diffusion Personalization (OPAD) [CVPR 2026 Findings] 🔵
    [Paper]
    Teacher-student adversarial distillation for personalizing one-step text-to-image diffusion models.

Image Super-Resolution

  • One-Step Effective Diffusion Network for Real-World Image Super-Resolution (OSEDiff) [NeurIPS 2024] 🔵
    [Paper] [Code]

  • One Step Diffusion-based Super-Resolution with Time-Aware Distillation (TAD-SR) [NeurIPS 2024] 🔵
    [Paper] [Code]

  • TSD-SR: One-Step Diffusion with Target Score Distillation [CVPR 2025] 🔵
    [Paper] [Code]

  • Consistency Trajectory Matching for One-Step Generative Super-Resolution [ICCV 2025] 🔵
    [Paper] [Code]

  • FlowSR: Fast Image Super-Resolution via Consistency Rectified Flow [ICCV 2025] 🟢
    [Paper]

  • One-Step Diffusion Transformer for Controllable Real-World Image SR (ODTSR) [CVPR 2026] 🔵
    [Paper] [Code]

  • TEASR: Training-efficient Any-step Diffusion Transformer for Real-World Image Super-Resolution [2026] 🔵
    [Paper]
    Self-adversarial any-step Real-ISR distillation in a single diffusion model, supporting both one-step and multi-step restoration.

Video Generation

  • SF-V: Single Forward Video Generation Model [NeurIPS 2024] 🔵
    [Paper] [Code]

  • OSV: One Step is Enough for High-Quality Image to Video Generation [CVPR 2025] 🔵
    [Paper]

  • Diffusion Adversarial Post-Training for One-Step Video Generation [ICML 2025] 🔵
    [Paper]

  • DOVE: Efficient One-Step Diffusion for Real-World Video Super-Resolution [NeurIPS 2025] 🔵
    [Paper] [Code]

  • One-Step Diffusion for Detail-Rich Video Super-Resolution (DLoRAL) [NeurIPS 2025] 🔵
    [Paper]

  • Autoregressive Adversarial Post-Training for Real-Time Interactive Video Generation [NeurIPS 2025] 🔵
    [Paper]

  • Motion Consistency Model (MCM) [NeurIPS 2024] 🔵
    [Paper] [Code]
    Single-phase video distillation with disentangled motion and appearance learning.

  • DCM: Dual-Expert Consistency Model for Video Generation [ICCV 2025] 🔵
    [Paper]
    Parameter-efficient dual-expert consistency decoupling semantics and details.

3D Generation

  • DIFIX3D+: Improving 3D Reconstructions with Single-Step Diffusion Models [CVPR 2025 Oral] 🔵
    [Paper] [Code]
    Single-step 3D reconstruction and novel view synthesis; CVPR 2025 Best Paper Finalist.

  • VideoScene: Distilling Video Diffusion Model to Generate 3D Scenes in One Step [CVPR 2025] 🔵
    [Paper] [Code]
    First one-step 3D scene generation by distilling video diffusion.

Robotics / Control

  • OneDP: Fast Visuomotor Policies via Diffusion Distillation [ICML 2025] 🔵
    [Paper]
    Distilling pretrained diffusion policies into single-step policies for robot control.

Other Applications

  • OSDFace: One-Step Diffusion Model for Face Restoration [CVPR 2025] 🔵
    [Paper] [Code]

  • Adjoint Matching: Fine-tuning Flow and Diffusion Generative Models [ICLR 2025] 🟢
    [Paper]

  • Reward Guided Latent Consistency Distillation (RG-LCD) [ICLR 2025] 🔵
    [Paper]

  • Diffusion2GAN: Distilling Diffusion Models into Conditional GANs [ECCV 2024] 🔵
    [Paper]

  • Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation [AAAI 2024] 🔵
    [Paper] [Code]

  • DKDM: Data-Free Knowledge Distillation for Diffusion Models with Any Architecture [CVPR 2025] 🔵
    [Paper] [Code]

Unified Theory

  • Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction [NeurIPS 2025]
    [Paper]
    A unified theoretical framework proving 10+ existing distillation methods are special cases of diffusion divergence minimization.

  • High-Order Flow Matching: Unified Framework [NeurIPS 2025]
    [Paper]
    A unified high-order flow-matching framework.

  • On the Design of One-step Diffusion via Shortcutting Flow Paths [ICLR 2026] 🟢
    [Paper] [Code]
    Theoretical analysis of supervision signal variance for designing one-step flow shortcuts.

Full Paper List

Type: 🟢 Native · 🔵 Distillation · 🟡 Hybrid
Category: CM = Consistency Models · FS = Flow Straightening · DDM = Distribution Divergence Min. · ADV = Adversarial Training · SC = Shortcut / Direct Mapping · DM = Drifting Models · APP = Applications · TH = Unified Theory · FW = Foundational Work

Date Paper Venue Type Category Links
2026.06 TEASR: Training-efficient Any-step Diffusion Transformer for Real-World Image Super-Resolution 2026 🔵 APP [Paper]
2026.05 W-Flow: One-Step Generative Modeling via Wasserstein Gradient Flows 2026 🟢 DM [Paper]
2026.05 FlowSR: Fast Image SR via Consistency Rectified Flow ICCV 2025 🟢 APP [Paper]
2026.04 SubFlow: Sub-mode Conditioned Flow Matching for Diverse One-Step Generation 2026 🟢 FS [Paper]
2026.03 Gradient Flow Drifting: Generative Modeling via Wasserstein Gradient Flows 2026 🟢 DM [Paper]
2026.03 Generative Drifting is Secretly Score Matching: a Spectral and Variational Perspective 2026 🟢 DM [Paper]
2026.03 WaDi: Weight Direction-aware Distillation for One-step Image Synthesis CVPR 2026 🔵 DDM [Paper] [Code]
2026.03 SoFlow: Solution Flow Models for One-Step Generative Modeling ICLR 2026 🟢 FS [Paper] [Code]
2026.02 PixelRush: Ultra-Fast, Training-Free High-Resolution Image Generation via One-step Diffusion CVPR 2026 🔵 APP [Paper]
2026.02 Bi-Anchor Interpolation Solver for Accelerating Generative Modeling ICML 2026 🔵 SC [Paper] [Code]
2026.02 Generative Modeling via Drifting 2026 🟢 DM [Paper] [Code]
2026.01 One-step Latent-free Image Generation with Pixel Mean Flows (pMF) 2026 🟢 FS [Paper]
2026.01 DiffRatio: Training One-Step Diffusion Models Without Teacher Supervision ICLR 2025 🟢 DDM [Paper] [Code]
2025.12 High-Order Flow Matching: Unified Framework NeurIPS 2025 — TH [Paper]
2025.12 TwinFlow: One-step Generation with Self-adversarial Flows ICLR 2026 🟢 ADV [Paper] [Code]
2025.12 On the Design of One-step Diffusion via Shortcutting Flow Paths ICLR 2026 🟢 TH [Paper] [Code]
2025.11 ODTSR: One-Step Diffusion Transformer for Controllable Real-World Image SR CVPR 2026 🔵 APP [Paper] [Code]
2025.10 Adversarial Concept Distillation for One-Step Diffusion Personalization (OPAD) CVPR 2026 Findings 🔵 APP [Paper]
2025.11 Consistency Trajectory Matching for One-Step Generative SR ICCV 2025 🔵 APP [Paper] [Code]
2025.11 DOVE: Efficient One-Step Diffusion for Real-World Video SR NeurIPS 2025 🔵 APP [Paper] [Code]
2025.10 One-Step Diffusion for Detail-Rich Video SR (DLoRAL) NeurIPS 2025 🔵 APP [Paper]
2025.10 Autoregressive Adversarial Post-Training for Real-Time Interactive Video Gen NeurIPS 2025 🔵 APP [Paper]
2025.10 Uni-Instruct: One-step Diffusion via Unified Diffusion Divergence Instruction NeurIPS 2025 — TH [Paper]
2025.10 How to Build a Consistency Model: Learning Flow Maps via Self-Distillation NeurIPS 2025 🟡 CM [Paper] [Code]
2025.10 Diffusion Adversarial Post-Training for One-Step Video Generation ICML 2025 🔵 APP [Paper]
2025.10 Any-step Generation via RCGM ICLR 2026 🟢 CM [Paper] [Code]
2025.09 SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation ICCV 2025 🔵 CM [Paper]
2025.07 Score-of-Mixture Training ICML 2025 Spotlight 🟢 DDM [Paper]
2025.06 Noise Consistency Training (NCT) NeurIPS 2025 🟢 SC [Paper] [Code]
2025.06 Align Your Flow: Scaling Continuous-Time Flow Map Distillation NeurIPS 2025 🔵 FS [Paper]
2025.06 Revisiting Diffusion Models: From Pre-training to One-Step Generation (D2O) ICML 2025 🟢 SC [Paper] [Code]
2025.06 Diff-Instruct++: Training One-step T2I Generator to Human-Preferred 2024 🔵 DDM [Paper] [Code]
2025.06 One Step Diffusion via Shortcut Models ICLR 2025 Oral 🟢 SC [Paper] [Code]
2025.06 DCM: Dual-Expert Consistency Model for Video ICCV 2025 🔵 APP [Paper]
2025.06 Variational Flow Matching (S-VFM) CVPR 2026 🟢 FS [Paper]
2025.05 TSD-SR: One-Step Diffusion with Target Score Distillation CVPR 2025 🔵 APP [Paper] [Code]
2025.05 Beyond Optimal Transport: Model-Aligned Coupling for Flow Matching CVPR 2026 Findings 🟢 FS [Paper] [Code]
2025.05 Mean Flows for One-step Generative Modeling (MeanFlow) NeurIPS 2025 🟢 FS [Paper] [Code]
2025.05 VideoScene: One-Step 3D Scene Generation CVPR 2025 🔵 APP [Paper] [Code]
2025.04 OSV: One Step is Enough for High-Quality Image to Video Generation CVPR 2025 🔵 APP [Paper]
2025.03 Learning Few-Step Diffusion Models by Trajectory Distribution Matching (TDM) ICCV 2025 🔵 FS [Paper] [Code]
2025.03 DIFIX3D+: Single-Step 3D Reconstruction CVPR 2025 Oral 🔵 APP [Paper] [Code]
2025.02 You Only Sample Once (YOSO) ICLR 2024 🟢 ADV [Paper] [Code]
2025.02 One-step Diffusion Models with f-Divergence Distribution Matching (f-distill) 2025 🔵 DDM [Paper]
2025.02 Inverse Flow and Consistency Models ICML 2025 🟢 CM [Paper]
2025.02 High-Order Matching for One-Step Shortcut Diffusion Models ICLR 2025 🟢 SC [Paper]
2025.02 Simple Distillation for One-Step Diffusion Models (CED) NeurIPS 2025 🔵 CM [Paper]
2025.02 End-to-End Single-Step Flow Matching via Direct Models (FlowFit) ICLR 2025 🟢 FS [Paper]
2024.12 Adversarial Score identity Distillation (SiDA) ICLR 2025 🔵 DDM [Paper] [Code]
2024.12 EM Distillation for One-step Diffusion Models NeurIPS 2024 🔵 DDM [Paper]
2024.12 Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps AAAI 2024 🔵 APP [Paper] [Code]
2024.11 Optimal Flow Matching: Learning Straight Trajectories in Just One Step NeurIPS 2024 🟢 FS [Paper] [Code]
2024.11 SwiftBrush: One-Step T2I Diffusion with Variational Score Distillation CVPR 2024 🔵 DDM [Paper] [Code]
2024.11 OSDFace: One-Step Diffusion Model for Face Restoration CVPR 2025 🔵 APP [Paper] [Code]
2024.10 Consistency Models Made Easy ICLR 2025 🟢 CM [Paper]
2024.10 One-Step Effective Diffusion Network for Real-World Image SR (OSEDiff) NeurIPS 2024 🔵 APP [Paper] [Code]
2024.10 SF-V: Single Forward Video Generation Model NeurIPS 2024 🔵 APP [Paper] [Code]
2024.10 Improving the Training of Rectified Flows NeurIPS 2024 🟢 FS [Paper] [Code]
2024.10 Reward Guided Latent Consistency Distillation (RG-LCD) ICLR 2025 🔵 APP [Paper]
2024.10 One-step Diffusion with Distribution Matching Distillation (DMD) CVPR 2024 🔵 DDM [Paper]
2024.10 One-Step Diffusion Distillation through Score Implicit Matching (SIM) NeurIPS 2024 🔵 DDM [Paper] [Code]
2024.10 OneDP: Fast Visuomotor Policies via Diffusion Distillation ICML 2025 🔵 APP [Paper]
2024.10 Di[M]O: Distilling Masked Diffusion into One-step ICCV 2025 🔵 SC [Paper] [Code]
2024.10 Truncated Consistency Models 2024 🟢 CM [Paper]
2024.09 Adjoint Matching: Fine-tuning Flow and Diffusion Generative Models ICLR 2025 🟢 APP [Paper]
2024.09 DKDM: Data-Free Knowledge Distillation for Diffusion Models CVPR 2025 🔵 APP [Paper] [Code]
2024.08 One Step Diffusion-based SR with Time-Aware Distillation (TAD-SR) NeurIPS 2024 🔵 APP [Paper] [Code]
2024.08 SwiftBrush v2: Make Your One-step Model Better Than Its Teacher ECCV 2024 🔵 DDM [Paper] [Code]
2024.07 Training Smaller One-Step Diffusion Models with Rectified Flow (SlimFlow) ECCV 2024 🔵 FS [Paper] [Code]
2024.06 Motion Consistency Model (MCM) NeurIPS 2024 🔵 APP [Paper] [Code]
2024.05 Improved Distribution Matching Distillation (DMD2) NeurIPS 2024 Oral 🔵 DDM [Paper] [Code]
2024.05 Diffusion2GAN: Distilling Diffusion Models into Conditional GANs ECCV 2024 🔵 APP [Paper]
2024.05 PeRFlow: Piecewise Rectified Flow NeurIPS 2024 🔵 FS [Paper] [Code]
2024.05 Phased Consistency Models (PCM) NeurIPS 2024 🔵 CM [Paper] [Code]
2024.04 Score identity Distillation (SiD) ICML 2024 🔵 DDM [Paper] [Code]
2024.04 Hyper-SD: Trajectory Segmented Consistency Model NeurIPS 2024 🔵 CM [Paper] [Code]
2024.03 Consistency Trajectory Models (CTM) ICLR 2024 🔵 CM [Paper] [Code]
2024.03 InstaFlow: One Step is Enough for High-Quality T2I Generation ICLR 2024 🔵 FS [Paper] [Code]
2024.03 Latent Adversarial Diffusion Distillation (LADD) ECCV 2024 🔵 ADV [Paper]
2024.02 SDXL-Lightning: Progressive Adversarial Diffusion Distillation 2024 🔵 ADV [Paper] [Code]
2024.01 Rectified Diffusion: Straightness Is Not Your Need ICLR 2025 🟢 FS [Code]
2023.11 Adversarial Diffusion Distillation (ADD) ECCV 2024 🔵 ADV [Paper]
2023.10 Improved Techniques for Training Consistency Models (iCT) ICLR 2024 🟢 CM [Paper]
2023.10 Latent Consistency Models (LCM) ICLR 2024 🔵 CM [Paper]
2023.06 BOOT: Data-free Distillation of Denoising Diffusion Models with Bootstrapping ICML 2023 🔵 DDM [Paper]
2023.05 Diff-Instruct: Universal Knowledge Transfer from Pre-trained Diffusion Models NeurIPS 2023 🔵 DDM [Paper]
2023.03 Consistency Models ICML 2023 🟡 CM [Paper]
2023.03 TRACT: Transitive Closure Time-Distillation 2023 🔵 FS [Paper] [Code]
2022.10 Flow Matching for Generative Modeling ICLR 2023 🟢 FW [Paper]
2022.09 Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow ICLR 2023 Spotlight 🟢 FW [Paper]
2022.02 Progressive Distillation for Fast Sampling of Diffusion Models ICLR 2023 🔵 FW [Paper]

Related Resources

Foundational Works

These papers are not one-step generation methods themselves, but provide the theoretical foundations upon which many one-step methods are built.

  • Progressive Distillation for Fast Sampling of Diffusion Models [ICLR 2023] 🔵
    [Paper]
    Iterative distillation that progressively halves the number of sampling steps; foundational for step-compression research.

  • Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow [ICLR 2023 Spotlight] 🟢
    [Paper]
    Learning ODEs that follow straight paths; the reflow procedure inspired many one-step flow methods.

  • Flow Matching for Generative Modeling [ICLR 2023] 🟢
    [Paper]
    Simulation-free training of continuous normalizing flows; the foundational framework behind MeanFlow, SoFlow, etc.

Contributing

Contributions are very welcome! Please read CONTRIBUTING.md for paper-entry formatting, category rules, and the native-vs-distillation labeling convention before submitting a pull request.

When adding a paper, classify it by its primary mathematical contribution (the principle that makes one-step generation work), then mark it 🟢 native, 🔵 distillation, or 🟡 hybrid, and place it chronologically within the category.

Citation

If you find this list useful in your research, please consider citing it:

@misc{awesome-one-step-generation,
  title        = {Awesome One-Step Generation},
  howpublished = {\url{https://github.com/Luciferbobo/Awesome-One-Step-Generation}},
  year         = {2026}
}

About

A curated list of papers, code, and resources for one-step diffusion models that turn noise into high-quality samples in a single neural network forward pass.

Resources

Readme
MIT license

Contributing

Contributing
Activity

Stars

172 stars

Watchers

0 watching

Forks

6 forks
Report repository

Releases

Packages

Contributors