One-Step Diffusion Models

Updated 6 April 2026

One-Step Diffusion is a generative modeling framework that condenses the traditional iterative denoising process into a single neural network pass by matching the distribution of multi-step teachers.
Training leverages techniques like direct distillation, score implicit matching, and GAN-based alignment to bridge fidelity gaps and ensure robust sample quality.
Applications span image synthesis, compression, face and video restoration, speech conversion, and robotics, achieving significant speedups and competitive performance benchmarks.

One-Step Diffusion

One-step diffusion refers to a class of generative models and sampling techniques in which the entire reverse diffusion process, typically implemented as an iterative chain of neural network evaluations, is collapsed into a single generative pass. Originally developed to address the computational inefficiency of multi-step denoising diffusion models (DDMs), one-step diffusion methods span applications in image synthesis, sequence modeling, compression, video restoration, robotics, and speech conversion. Recent research demonstrates that with appropriate training protocols, theoretical frameworks, and architectural adaptations, one-step diffusion models can achieve sample quality and perceptual realism that rival, or in certain benchmarks exceed, their multi-step diffusion teacher models.

1. Theoretical Foundations and Distillation Objectives

Traditional denoising diffusion models generate data by numerically inverting a forward (often Gaussian) SDE/ODE via many sequential denoising steps. One-step diffusion methods challenge this paradigm by seeking a mapping from noise to data (or conditional data) that produces high-fidelity samples in a single neural network evaluation. The central challenge is bridging the distributional gap between the one-step generator and the original diffusion process, which occupies different optimization basins and exploits different inductive biases.

Various theoretical objectives have emerged to achieve distribution-level alignment between the student (one-step generator) and the teacher (multi-step diffusion). Notable frameworks include:

Distribution Matching Distillation (DMD): Minimizes a distributional KL divergence between the one-step generator and teacher via score function differences in noise-augmented space (Yin et al., 2023).
Score Implicit Matching (SIM): Establishes a data-free loss based on the difference between marginal score functions of the student and teacher, leveraging a score-gradient theorem to provide tractable gradients for generator optimization (Luo et al., 2024).
Unified f-divergence Expansion (Uni-Instruct): Derives an expanded f-divergence integral along the diffusion trajectory, unifying KL-based and score-based objectives. The parameter gradient yields tractable combinations of regression and score-matching terms, with parametrizable divergence for mode-seeking or mode-covering behavior (Wang et al., 27 May 2025).
Bregman Density Ratio Matching (Di-Bregman): Formulates distillation as convex Bregman divergence minimization over density ratios between student and teacher, showing that many existing objectives (reverse-KL, least-squares, etc.) are special cases (Zhu et al., 19 Oct 2025).
One-step Shortcut Models: Direct parameterization of the flow map or velocity fields, trained with a self-consistency loss to collapse multiple-step denoising into a single step, often without explicit teacher-student distillation (Frans et al., 2024, Lin et al., 3 Dec 2025).

Empirically, instance-based regression to teacher outputs is often ineffective at single-step. Frameworks that prioritize distribution-level alignment—sometimes incorporating auxiliary GAN losses for realism—have demonstrated robust improvements in one-step model fidelity and stability (Zheng et al., 2024).

2. Training Algorithms and Network Architectures

Training procedures for one-step diffusion models exhibit substantial diversity:

Direct Distillation (One-step Student): The student generator $G_\theta(z)$ is trained via implicit score matching, distribution matching, or f-divergence expansion to minimize discrepancy with the teacher's data distribution.
GAN-based Distributional Alignment: A discriminator is adversarially trained to distinguish real data from student samples, encouraging the one-step network to match the teacher or data distribution at a global level (Zheng et al., 2024).
Self-Consistency and Shortcuts: Networks are conditioned on time and step size, and trained via self-consistency losses to ensure that single large steps are consistent with compositions of multiple small steps (shortcutting the flow path) (Frans et al., 2024, Lin et al., 3 Dec 2025).
Equilibrium Models: Deep equilibrium transformers (DEQ) are employed as the student, leveraging their implicit infinite-depth representation to more flexibly approximate complex one-step flows (Geng et al., 2023).
Plug-in Velocity and EMA: Advanced training recipes, such as plug-in estimation of marginal velocity and class-consistent batching, are used to stabilize variance and accelerate convergence in shortcut flow-based approaches (Lin et al., 3 Dec 2025).

Architecturally, one-step diffusion models are implemented using U-Nets (often with frozen convolutional backbones and trainable normalization layers), transformers (DiT, SiT, Equilibrium Transformer), and task-appropriate modules (e.g., VAE for image compression, content encoders for voice conversion, VQVAE-style visual representation embedders for face restoration).

3. Mathematical Formulation and Closed-Form Sampling

The key innovation in one-step diffusion models is the identification of analytic or learnable mappings that reverse the noising process in a single step. For models built on Gaussian diffusion SDEs/ODEs, the reverse step typically takes the form:

$\hat x_0 = \frac{x_t - \sqrt{1-\bar\alpha_t}\,\epsilon_\theta(x_t, c, t)}{\sqrt{\bar\alpha_t}}$

where $x_t$ is the noised input (possibly at a pseudo-timestep reflecting modulation or compression rate), $\epsilon_\theta$ is the learned score/denoising network, $c$ denotes conditional input or prompts, and $\bar\alpha_t$ is the cumulative noise schedule (Jia et al., 2 Feb 2026, Guo et al., 22 May 2025, Wang et al., 2024). In shortcut models and flow-matching frameworks, the one-step generator is parameterized to map from pure noise directly to data by integrating a velocity field or flow map conditioned on the entire step size (Frans et al., 2024, Lin et al., 3 Dec 2025).

Deterministic importance weighting and volume consistency regularization are sometimes employed in probabilistic inference settings, allowing for both accurate sampling and robust evidence estimation (Jutras-Dube et al., 4 Dec 2025).

4. Applications and Domain-Specific Variants

One-step diffusion models have been adopted and demonstrated in multiple domains:

Image Generation and Editing: Unconditional/class-conditional image synthesis (CIFAR-10, FFHQ, AFHQ, ImageNet) and editing tasks are performed with only a single forward evaluation, attaining FID scores that rival or surpass multi-step teachers (Zheng et al., 2024, Wang et al., 27 May 2025).
Perceptual Image Compression: One-step diffusion enables high-fidelity, low-latency generative codecs (e.g., OneDC, OSDiff, OSCAR), decoupling compression and semantic guidance, and achieving up to 46× faster decoding at reduced bitrates (Xue et al., 22 May 2025, Jia et al., 2 Feb 2026, Guo et al., 22 May 2025).
Face and Video Restoration: OSDFace couples a visual prompt from a VQVAE-based embedder with adversarial-guided one-step latent denoising for identity-preserving face restoration (Wang et al., 2024). For video snapshot compressive imaging, one-step diffusion inverts mean-reverting SDEs matching hardware modulation, drastically accelerating video reconstruction (Wang et al., 19 Dec 2025).
Speech Processing: In one-step VC models (FastVoiceGrad, FasterVoiceGrad), adversarial and score-based distillation transfer a multi-step voice conversion model into a fast single-step student, maintaining both quality and speaker similarity (Kaneko et al., 2024, Kaneko et al., 25 Aug 2025).
Robotics and Policy Learning: The One-Step Diffusion Policy distills multi-step visuomotor diffusion policies into a single-step action generator, enabling real-time control rates (>60 Hz) in manipulation and imitation learning settings (Wang et al., 2024).
Language Modeling: DLM-One shows that entire text sequences can be generated in one step in embedding space by distilling a continuous DLM via score alignment, reducing inference complexity by up to 500× (Chen et al., 30 May 2025).

5. Empirical Performance and Comparative Results

Recent one-step diffusion frameworks set new SOTA results in various settings:

Model/Framework	Benchmarks	Domain	Best FID	Key Speedup
Uni-Instruct (Wang et al., 27 May 2025)	CIFAR-10, ImageNet-64	Image synthesis	1.46 (CIFAR-10), 1.02 (ImageNet-64)	>50× faster†
GDD-I (Zheng et al., 2024)	CIFAR-10, FFHQ, etc.	Image synthesis	1.54 (CIFAR-10), 0.85 (FFHQ)	35–79×
DMD (Yin et al., 2023)	ImageNet-64	Image synthesis	2.62 (vs. 2.32 multi-step)	512×
MeanFlow/ESC (Lin et al., 3 Dec 2025)	ImageNet-256 (CFG)	Image synthesis	2.85 FID50k (from scratch)	>100×
OSDFace (Wang et al., 2024)	CelebA-Test, LFW	Face restoration	17.06 (FID HQ), 45.42 (FFHQ)	250–1000×
OSDiff (Jia et al., 2 Feb 2026)	Kodak, CLIC_2020	Compression	Rate–distortion parity w/ multistep	46×
OSCAR (Guo et al., 22 May 2025)	CLIC, DIV2K, Kodak	Compression	−8–10% BD-rate over prior SOTA	20–50×
DLM-One (Chen et al., 30 May 2025)	QQP, Quasar-T, Wiki	Text generation	≤1% degradation vs. teacher BLEU	500×
3One2/RegDif (Wang et al., 19 Dec 2025)	SCI, videos	Video reconstruction	∼2 dB PSNR gain vs. strong baseline	50–200×

†Speedup is in terms of wall-clock or MACs compared to multi-step diffusion or autoregressive baselines.

Leading models exhibit high perceptual quality (LPIPS, DISTS), robust sample diversity, and, in compression, significant bitrate or latency reductions. Short-cut and rectified flow innovations enable from-scratch one-step models that compete with multi-step or distilled alternatives (Frans et al., 2024, Zhu et al., 2024, Lin et al., 3 Dec 2025).

6. Limitations, Extensions, and Future Directions

Despite their empirical success, one-step diffusion models face open challenges:

Fidelity Gaps: While recent FIDs are competitive, a small gap remains vs. the strongest multi-step samplers, especially in the rendering of fine details, complex scenes, or extremal degradation (ultra-low bitrates) (Guo et al., 22 May 2025, Luo et al., 2024).
Training Instability: Distribution-matching GAN losses or density-ratio estimation (for advanced divergences) can introduce instability; careful architecture and loss weighting are critical (Wang et al., 27 May 2025, Zhu et al., 19 Oct 2025).
Generalization Limits: In some domains, such as text-to-image and 3D generation, one-step models may underperform on out-of-distribution prompts or extremely long sequences. Progressive or adaptive-step inference may improve trade-offs (Luo et al., 2024, Chen et al., 30 May 2025).
Resource Requirements: Rich hyperparameter tuning, vast teacher-generated synthetic datasets, and capacity-matched architectures are often needed for SOTA performance, which may not scale efficiently to higher resolutions or new modalities (Roy et al., 3 Jul 2025).
Theory-Practice Gap: Although f-divergence unification and flow-matching offer rigorous theoretical grounding, practical implementation may depart from ideal settings (e.g., non-Gaussian noise, learned schedules, non-canonical ODEs). Further theoretical development is ongoing (Wang et al., 27 May 2025, Lin et al., 3 Dec 2025, Jutras-Dube et al., 4 Dec 2025).

Current extensions under investigation include hybrid multi-step refinement, better ratio estimation, on-the-fly guidance adaptation, and scaling up to video, 3D, and multimodal generation.

7. Summary Table of Key Frameworks

Method	Theoretical Foundation	Achieves SOTA?	Core Domain(s)	Key Cited Papers
Uni-Instruct	Diffusion f-divergence expansion	Yes (CIFAR, ImageNet)	Images, Text-to-3D	(Wang et al., 27 May 2025)
SIM	Score implicit matching	Yes (CIFAR, T2I)	Images, Text-to-Image	(Luo et al., 2024)
DMD	Distribution-matching KL	Yes (ImageNet)	Images	(Yin et al., 2023)
GDD-I	Distributional GAN distillation	Yes (CIFAR, FFHQ)	Images	(Zheng et al., 2024)
Shortcut/ESC	DDIM/flow-matching shortcut	Yes (ImageNet-256)	Images (from scratch)	(Lin et al., 3 Dec 2025, Frans et al., 2024)
DLM-One	Score distillation (embedding)	Yes (Seq2seq)	Language generation	(Chen et al., 30 May 2025)
SlimFlow	Rectified-flow + model compression	Yes (CIFAR, FFHQ, IN64)	Images (small models)	(Zhu et al., 2024)
OSDiff/OSCAR/OneDC	Closed-form one-step denoising	Yes (compression)	Image/video compression	(Jia et al., 2 Feb 2026, Guo et al., 22 May 2025)
OneDP (robotics)	KL chain distillation	Yes (simulation/real)	Robotics, Visuomotor policy	(Wang et al., 2024)
OSDFace/RegDif	One-step denoising + prompt/score	Yes (restoration/SCI)	Face restoration, video SCI	(Wang et al., 2024, Wang et al., 19 Dec 2025)
FastVoiceGrad	ACDD/ADCD distilled VC	Yes (VCTK, LibriTTS)	Voice conversion	(Kaneko et al., 2024, Kaneko et al., 25 Aug 2025)

One-step diffusion now constitutes a general and high-impact framework for fast, distribution-matched generative modeling across vision, language, audio, and control domains. Its practical adoption is driven by theoretical advances in divergence minimization, shortcut mechanisms, data-free distillation, and stability-enhancing architectural design.