One-Step Diffusion Models

• One-step diffusion is a generative modeling paradigm that compresses multi-step denoising into one efficient neural network evaluation through specialized distillation and loss functions.
• It employs diverse strategies like f-divergence expansion and architectural innovations such as low-rank adaptations and deep equilibrium models to preserve high-frequency and semantic details.
• State-of-the-art implementations demonstrate high fidelity in image, video, speech synthesis, and compression tasks while significantly reducing inference latency.

One-step diffusion is a paradigm in generative modeling where the iterative (typically multi-step) denoising trajectory of diffusion models is “compressed” into a single neural network evaluation. Originally motivated by the need for rapid and resource-efficient generation, this concept has now matured with rigorous theoretical frameworks and state-of-the-art empirical results across domains such as image, language, audio, and video synthesis. One-step diffusion encompasses model construction, specialized distillation strategies, architectural modifications, and regularizations, all engineered to enable high-fidelity data synthesis with only a single forward pass.

1. Motivation and General Principles

Conventional diffusion models produce samples by reversing a noise-adding process in hundreds or thousands of small steps, with each step progressively denoising the current state. While this yields high sample quality, it severely limits applicability in latency-sensitive or resource-restricted deployments. The core principle of one-step diffusion is to approximate (or “distill”) the denoising path of a multi-step process into a single operation, mapping a noise vector (or compressed conditional input) directly to a high-quality sample. Several research directions have established that this is feasible given (1) carefully constructed distillation losses, (2) architectures capable of expressing the requisite conditional mapping, and (3) strategies for preserving critical high-frequency and semantic details.

One-step diffusion distinguishes itself from standard model acceleration techniques (which merely reduce step count) by fundamentally altering the training objective so as to directly produce the final sample distribution in a single evaluation. Recent frameworks formalize this process as matching not just output images, but the distributional, score-based, or divergence measures along entire diffusion chains.

2. Distillation Strategies and Theoretical Formulations

The transition from multi-step to one-step generative diffusion generally proceeds via network distillation—training a “student” generator to mimic the output of a full diffusion process defined by a “teacher” model. Early approaches relied on progressive or consistency-based distillation, but these typically yielded subpar quality or required resource-intensive multi-student architectures.

Recent advances unify distillation methods under expanded f-divergence frameworks, where the difference between teacher and student distributions is quantified as an integrated divergence over the diffusion trajectory. Uni-Instruct (Wang et al., 27 May 2025), for example, shows:

$$\mathcal{D}_f(q_0 \parallel p_\theta) = \int_0^T \frac{1}{2}g^2(t)\, \mathbb{E}_{p_{\theta,t}}\left[ \left( \frac{q_t}{p_{\theta,t}} \right)^2 f''\left(\frac{q_t}{p_{\theta,t}}\right) \|\mathbf{s}_{p_{\theta,t}}(\mathbf{x}_t) - \mathbf{s}_{q_t}(\mathbf{x}_t)\|_2^2 \right]dt$$

This construction unifies prior KL-based, score-matching, and adversarial strategies by yielding tractable surrogate gradients for the otherwise intractable distributions induced by one-step generators (see also Score Implicit Matching (Luo et al., 22 Oct 2024) and Distribution Matching Distillation (Yin et al., 2023)). These approaches enable direct score alignment between the one-step mapping and the diffusion process, giving rise to theoretically principled and empirically superior outcomes.

The table below summarizes key loss types employed in modern one-step diffusion training:

Distillation Strategy	Objective	Application Domain
f-divergence expansion	Integral over diffusion time (score function matching)	Images, text-to-3D
KL divergence gradient	Difference of teacher and student scores	Images, policies
Distributional GAN	Batch-level adversarial statistics	Images, video, speech
Regression/perceptual	Instance-level image or feature matching (e.g. LPIPS)	Super-resolution, VQ

3. Architectural Approaches and Data-Dependent Design

One-step diffusion models adapt both the backbone and conditioning interface to enable rich, context-sensitive mappings in one evaluation. Several representative approaches include:

• Low-rank Adaptation (LoRA-based): HiPA (Zhang et al., 2023) inserts low-rank adaptation modules to specifically boost high-frequency capabilities, shown to be lacking in collapsed one-step outputs.
• Deep Equilibrium Models: GET (Geng et al., 2023) implements the generator as the fixed point of a weight-tied transformer block (DEQ). This implicit depth-wise recursion is crucial for compressing complex denoising paths.
• Mixture-of-Experts: MSD (Song et al., 30 Oct 2024) partitions the condition or class space among several distilled students, vastly improving both quality and computational efficiency for high-cardinality conditions.
• Latent Consistency and Cross-Attention: OSDFace (Wang et al., 26 Nov 2024) and SNOOPI (Nguyen et al., 3 Dec 2024) introduce vector-quantized prompts and cross-attention steering to address fine-grained control in restoration or text-to-image synthesis.
• Conditional Step Size Parameterization: Shortcut models (Frans et al., 16 Oct 2024) condition not only on the noise level but explicitly on the desired step size, allowing seamless transition between one and many-step regimes.

Efficient architectures are frequently leveraged, such as reusing large teacher UNets for flow distillation (FluxSR (Li et al., 4 Feb 2025)), or compressing models via SlimFlow (Zhu et al., 17 Jul 2024) for edge deployment. Some models, e.g., FasterVoiceGrad (Kaneko et al., 25 Aug 2025), further accelerate inference by simultaneously distilling heavy content encoders into lightweight, one-step CNNs.

4. Domain-Specific Techniques and Applications

One-step diffusion methods have been engineered for diverse domains and tasks:

• Text-to-Image Synthesis: HiPA (Zhang et al., 2023), SNOOPI (Nguyen et al., 3 Dec 2024), and SIM (Luo et al., 22 Oct 2024) develop methods to recover critical high-frequency and semantic details, ensuring realism and prompt-faithfulness.
• Super-Resolution and Deblurring: FluxSR (Li et al., 4 Feb 2025) leverages flow trajectory distillation and specialized regularizations (TV-LPIPS, ADL) to preserve high-frequency textures while avoiding artifacts. OSDD (Liu et al., 9 Mar 2025) demonstrates the integration of synthetic datasets and dynamic adapters for robustness in motion deblurring.
• Image Compression: OSCAR (Guo et al., 22 May 2025) and OneDC (Xue et al., 22 May 2025) unify multi-rate latent quantization and one-step denoising frameworks, employing semantic distillation and hybrid optimization to achieve state-of-the-art rate-distortion and perceptual scores across bitrates.
• Real-time Video Synthesis: OSA-LCM (Guo et al., 18 Dec 2024) introduces latent consistency models and adversarial motion-aware discriminators for one-step expressive portrait video generation.
• Voice Conversion: FastVoiceGrad (Kaneko et al., 3 Sep 2024) and FasterVoiceGrad (Kaneko et al., 25 Aug 2025) adapt adversarial and score-based distillation with optimized initial state handling, yielding high-quality speech and efficient operation.
• Behavioral Policy Distillation for Robotics: One-Step Diffusion Policy (Wang et al., 28 Oct 2024) achieves order-of-magnitude improvements in policy latency for real-world tasks by aligning the student and teacher distributions via KL-based score matching.

5. Performance Benchmarks and Comparison

State-of-the-art one-step models report major improvements in both sample quality and computational efficiency. On benchmark image datasets:

• Uni-Instruct (Wang et al., 27 May 2025) achieves an unconditional FID of 1.46 (conditional: 1.38) on CIFAR10 and 1.02 on ImageNet-64×64 with one-step, improving over its 79-step teacher.
• MSD (Song et al., 30 Oct 2024) reaches FID 1.20 on ImageNet-64×64 and 8.20 on zero-shot COCO2014 with smaller, specialized students.
• FluxSR (Li et al., 4 Feb 2025) demonstrates visual realism and competitive metrics (MUSIQ, MANIQA, TOPIQ) in high-fidelity super-resolution, while suppressing periodic artifacts through ADL and TV-LPIPS.
• OSDD (Liu et al., 9 Mar 2025) matches or surpasses multi-step transformer and CNN baselines in deblurring, with superior no-reference perceptual metrics and inference speed.
• OSCAR (Guo et al., 22 May 2025) and OneDC (Xue et al., 22 May 2025) set new standards in rate-distortion for high-quality image compression, with 20x speedup and over 40% bitrate reduction compared to prior methods.
• DLM-One (Chen et al., 30 May 2025) supports full-sequence language generation in one step, achieving a ~500× inference speedup with minimal loss in BLEU and ROUGE-L relative to its continuous diffusion teacher.

These results confirm that task-specific, theoretically-grounded one-step diffusion training can match or even exceed traditional multi-step sampling—resolving a long-standing speed-fidelity trade-off in diffusion-based generative modeling.

6. Technical Innovations and Limitations

Several distinctive technical strategies have emerged:

• Distributional Supervision: GDD (Zheng et al., 31 May 2024) and SIM (Luo et al., 22 Oct 2024) stress the efficacy of distributional (GAN-based or score-based) rather than instance-level losses.
• Layer Freezing and Specialization: Freezing most convolutional layers while fine-tuning only group norms and skips (GDD-I (Zheng et al., 31 May 2024)) exploits the temporal specialization in standard diffusion models and enhances one-step capability.
• Guidance and Prompting: SNOOPI (Nguyen et al., 3 Dec 2024) introduces random-scale classifier-free guidance and NASA (Negative-Away Steer Attention) to stabilize performance and allow explicit negative prompting for controllable generation.
• Efficient Adversarial Distillation: FastVoiceGrad (Kaneko et al., 3 Sep 2024) and FasterVoiceGrad (Kaneko et al., 25 Aug 2025) demonstrate that adversarial and score-based distillation losses, when performed directly in the conversion process, prevent identity mapping collapse and improve quality in VC tasks.
• Semantic Distillation for Compression: OneDC (Xue et al., 22 May 2025) transfers semantic knowledge from pretrained tokenizers into the hyperprior, overcoming the weaknesses of text-prompt-based guidance for complex images.

Despite these advances, one-step diffusion models are still bounded by the capacity of their architectures, the efficacy of distributional alignment, and, in some cases, residual artifacts (such as periodic textures in super-resolution). The quality of distilled models depends critically on teacher model expressiveness and the precision of the loss alignment during training.

7. Future Prospects and Implications

One-step diffusion is poised for broad impact across generative disciplines:

• Scalability: Techniques such as mixture-of-experts distillation, score matching expansions, and hybrid adversarial-score objectives are likely to generalize to higher-resolution, multi-modal, and long-horizon generative tasks.
• Real-time and Edge Applications: By dramatically reducing inference steps and model sizes (e.g., SlimFlow (Zhu et al., 17 Jul 2024)), these models become well-suited for deployment in mobile, embedded, and low-latency settings.
• Unified Theory: The emergence of unified frameworks such as Uni-Instruct (Wang et al., 27 May 2025) and the use of tractable f-divergence expansions provide a principled basis for analyzing and extending one-step distillation techniques, potentially leading to automated or adaptive curriculum design for knowledge transfer.
• Transfer to Non-image Modalities: LLMing with DLM-One (Chen et al., 30 May 2025), text-to-3D generation, and voice conversion validate the extension of one-step diffusion’s principles beyond vision, opening the door to cross-domain applications in interactive media, robotics, and content creation.

The progression from stepwise denoising to efficient one-step mapping—anchored in theoretical developments and diverse engineering innovations—marks a significant threshold in the practical deployment and fundamental understanding of diffusion-based generative models.