One-Step Diffusion Prior

• One-Step Diffusion Prior is a generative modeling technique that compresses multi-step diffusion processes into a single neural mapping for rapid sample generation.
• It leverages EM distillation and unified f-divergence frameworks to closely approximate complex data distributions while reducing computational overhead.
• Applications include image synthesis, super-resolution, and inverse problems, demonstrating significant throughput improvements over traditional diffusion models.

A one-step diffusion prior is a generative modeling paradigm wherein the expressive capacity of a multi-step diffusion model is compressed or distilled into a single neural network mapping, enabling fast, non-iterative sample generation while approximating the complex distributional behaviors of iterative diffusion. Adoption of one-step diffusion priors is motivated by both the computational inefficiency of traditional diffusion models—which require dozens to hundreds of network evaluations per sample—and theoretical insights indicating that, under suitable learning objectives and network parameterizations, a single-step mapping can inherit the generative coverage and perceptual fidelity of its multi-step teacher across a broad class of tasks.

1. Theoretical Motivation and Objective Formulations

The central objective in one-step diffusion prior research is to approximate the data marginal induced by a multi-step diffusion model using a direct, parameterized mapping. Early approaches such as EM Distillation (Xie et al., 2024) recast the problem as maximum-likelihood fitting of a latent-variable model, training a neural network generator $g_\theta(z)$ such that $x = g_\theta(z)$ with $z \sim \mathcal N(0, I)$ has a marginal matching the target diffusion prior. The expectation-maximization (EM) framework is used to express the log-likelihood gradient, involving joint samples from the product of the pretrained diffusion marginal $q_t(x)$ and the conditional posterior $p_t(z|x;\theta)$. The EM gradient is estimated by:

$$\nabla_\theta \mathcal{L}(\theta) = \mathbb{E}_{q_t(x) p_t(z|x;\theta)}[\nabla_\theta \log p_t(x|z;\theta)]$$

where $p_t(x|z;\theta)$ is typically parameterized as a Gaussian centered at $\alpha_t g_\theta(z)$ with variance $\sigma_t^2 I$.

Recent unifying theoretical advances, notably Uni-Instruct (Wang et al., 27 May 2025), frame all existing one-step approaches as minimizers of an expanded $f$-divergence between data and model distributions, expressed as an integral over the diffusion time axis. The diffusion expansion theorem shows:

$x = g_\theta(z)$0

Thus, various one-step loss functions (e.g., KL, score-based, adversarial) emerge as special cases of this general expansion.

2. Algorithms for One-Step Prior Training

Most one-step priors are trained either by distillation from a fixed teacher diffusion model or by "shortcutting" the diffusion process altogether via from-scratch training.

Distillation Frameworks

• EM Distillation (EMD) (Xie et al., 2024):
  • Alternates between E-step (joint Langevin MCMC over latent $x = g_\theta(z)$1 and reparameterization noise $x = g_\theta(z)$2 within $x = g_\theta(z)$3 coordinates) and M-step (gradient ascent on maximum-likelihood using joint samples).
  • Stabilizes updates using a noise-cancellation trick: tracking and subtracting the accumulated Gaussian noise injected during Langevin steps, which preserves unbiasedness while minimizing gradient variance.
  • Recovers mode-seeking (reverse-KL) variational score distillation in the $x = g_\theta(z)$4 joint-Langevin regime; multi-Langevin ($x = g_\theta(z)$5) interpolates toward forward-KL, covering more modes.
• Consistency and Knowledge Distillation:
  • Student networks are trained to map from a single, appropriately noised input (drawn from the forward diffusion process) directly to the data distribution, using the teacher's multi-step result as the ground truth. Losses typically employ $x = g_\theta(z)$6 or hybrid perceptual measures (Kawai et al., 18 Jul 2025).
• GAN Fine-Tuning (Zheng et al., 11 Jun 2025):
  • Diffusion models are pretrained as score-based estimators; then GAN objectives (e.g., non-saturating loss, multi-scale Projected GAN discriminator) are applied, converting the denoiser into a one-step generator. Empirically, freezing the majority of backbone parameters (e.g., 85%) during fine-tuning maintains stability and efficiency.
• Unified Divergence Objectives (Wang et al., 27 May 2025):
  • By expanding $x = g_\theta(z)$7-divergences along the diffusion trajectory, all major one-step training methods—including diffusion instruction (Diff-Instruct), $x = g_\theta(z)$8-distill, SIM/SiD, etc.—are shown to share the same mathematical substrate. A practical surrogate for the intractable diffusion expansion is derived, combining "DI" (density ratio-weighted score difference) and "SIM" (score inner product) losses with a GAN-based density ratio estimator.

Training from Scratch: Shortcut Models

• Shortcutting Flow Paths (Lin et al., 3 Dec 2025):
  • One-step samplers are learned directly by predicting the flow map from prior noise to data in a single step, regressing against a two-step "target" computed using numerical solvers (e.g., DDIM updates along a probability-flow ODE).
  • Key architectural choices are disentangled: parameterization (instantaneous vs. average velocity), time samplers, and training metrics (e.g., LPIPS).
  • Plug-in velocity (batch-level marginal velocity estimation), class-consistent batching (for classifier-free guidance), and gradual time-sampler scheduling yield statistically robust local velocity estimates and improved guidance; these techniques collectively define the ESC (Explicit & Easier Shortcut) model.

3. Noise Reparameterization, Latent Conditioning, and Guidance

Reparameterized sampling improves the stability and efficiency of joint Langevin and consistency-based methods. Expressing denoising in $x = g_\theta(z)$9 space for EM distillation and related methods allows for unified step-size scheduling and leverages learned student scores for $z \sim \mathcal N(0, I)$0.

For conditional tasks, semantic information is injected in various forms:

• Latent Compression Guidance (Xue et al., 22 May 2025):
  • OneDC employs a hyperprior extracted from quantized analysis transforms as a semantic guide, which is further distilled from a pretrained generative tokenizer to provide high-fidelity context to the one-step diffusion decoder.
• Data-Dependent Adaptation (Zhang et al., 2024):
  • Degradation-aware LoRA modules tailor model parameters to the input's estimated degradation during super-resolution, facilitating efficient adaptation while preserving the generative prior.
• Classifier-Free Guidance:
  • One-step priors often blend positive and negative attribute-conditioned forward passes (via classifier-free guidance) during inference to sharpen outputs and repel low-quality generations (Zhang et al., 2024).

4. Empirical Evaluation and Comparative Results

Recent works consistently demonstrate that one-step diffusion priors approach, match, or surpass the multi-step teacher in perceptual and distributional metrics, while incurring far lower computational costs.

Method	Application	Multi-Step FID	One-Step FID	Throughput Improvement	Reference
EM Distillation	ImageNet-64	1.43 (256 steps)	2.20 (EMD-16)	$z \sim \mathcal N(0, I)$1100$z \sim \mathcal N(0, I)$2	(Xie et al., 2024)
Uni-Instruct	ImageNet-64	2.35 (79 steps)	1.02	$z \sim \mathcal N(0, I)$379$z \sim \mathcal N(0, I)$4	(Wang et al., 27 May 2025)
OneDC	MS-COCO image compression	123.2 (multi-step)	69.7 (one-step)	20$z \sim \mathcal N(0, I)$5	(Xue et al., 22 May 2025)
ESC	ImageNet-256	--	2.85	--	(Lin et al., 3 Dec 2025)

Ablation studies underline that plug-in velocity estimation, noise-cancellation, and hybrid pixel/latent-domain losses are critical for performance, both in terms of FID and training stability. On domain-specific tasks, one-step models attain comparable perceptual scores (e.g., LPIPS, DISTS) and PSNR as their multi-step teacher or GAN-based peers, often outperforming previous single-step or consistency baselines.

5. Extended Applications and Generalizations

One-step diffusion priors are applicable beyond unconditional image generation:

• Posterior Amortization (Mammadov et al., 2024):
  • Conditional normalizing flows are trained to match diffusion model-induced posteriors, enabling one-step Bayesian sampling for inverse problems (e.g., denoising, inpainting) on both Euclidean and manifold data.
  • Experimental results on CelebA denoising and geospatial imputation indicate four orders of magnitude speedup relative to traditional diffusion solvers, with only modest loss in PSNR and SSIM.
• Super-Resolution (Zhang et al., 2024, Kawai et al., 18 Jul 2025):
  • One-step adaptation of large diffusion priors (with LoRA or knowledge distillation) delivers state-of-the-art perceptual and distortion metrics at high efficiency, handling burst and single-image inputs.
• Image Compression (Xue et al., 22 May 2025):
  • Semantic distillation of diffusion priors into one-step decoders yields drastic bitrate reductions and increased throughput, supporting both fidelity and realism at ultra-low bandwidth.
• Physics-Based and 3D GAN Priors (Zhu et al., 24 Jul 2025, Wang et al., 27 May 2025):
  • BokehDiff demonstrates one-step neural lens-blur rendering with physics-inspired attention, while Uni-Instruct extends the one-step prior concept to text-to-3D synthesis via NeRF optimization.

6. Limitations, Open Challenges, and Future Prospects

While one-step diffusion priors have closed the gap with multi-step models on many metrics, challenges remain:

• Reverse-KL or mode-seeking losses can lead to oversampling high-density regions while under-covering low-probability modes (reduced recall), especially in naively compressed/shortcut settings.
• The training time for highly accurate one-step priors is often nontrivial, particularly when relying on heavy multi-Langevin or adversarial schemes.
• Some domain-specific tasks may suffer from overfitting or require sophisticated conditioning to match the diversity of multi-step approaches.
• Theoretical understanding of the limits of expressivity and mode-coverage remains active, particularly regarding f-divergence scheduling and the role of density ratio estimation.

Advances in unified theoretical frameworks (Wang et al., 27 May 2025), modular shortcut design (Lin et al., 3 Dec 2025), and robust empirical strategies (e.g., plug-in velocity, semantic distillation, hybrid inference) are rapidly narrowing these gaps. Ongoing research is likely to focus on further reducing training overhead, improving diversity and theoretical guarantees, and extending one-step priors to high-resolution, multimodal, and structured data synthesis.