Diffusion Once and Done (DOD)

• Diffusion Once and Done (DOD) is a one-step diffusion paradigm that replaces iterative denoising with a single computed evaluation to achieve high task-specific performance.
• It leverages techniques such as adversarial matching, shortcutting probability flow paths, and task-specific distillation to ensure fidelity and efficiency.
• DOD spans diverse applications—including image synthesis, image restoration, anomaly detection, and kinodynamic planning—highlighting its versatile inference-time benefits.

Diffusion Once and Done (DOD) denotes a one-step diffusion paradigm in which a model performs the task-relevant diffusion computation once and terminates, rather than executing the multi-step iterative denoising chain characteristic of standard diffusion models. The term appears explicitly in "Diffusion Once and Done: Degradation-Aware LoRA for Efficient All-in-One Image Restoration" (Tang et al., 5 Aug 2025), but closely related formulations include "You Forward Once" for text-to-image synthesis (xu et al., 2023), "You Only Sample Once" for self-cooperative diffusion GANs (Luo et al., 2024), one-step diffusion by shortcutting probability flow paths (Lin et al., 3 Dec 2025), one-step diffusion scoring in inverse residual field space for unsupervised industrial anomaly detection (Zhang et al., 20 Apr 2026), and Train-Once Plan-Anywhere kinodynamic planning with diffusion policies used as informed samplers inside sampling-based planners (Hassidof et al., 28 Aug 2025). Across these works, DOD names a design objective: replacing repeated diffusion iterations with a single learned evaluation while preserving task-specific desiderata such as fidelity, restoration quality, anomaly separability, or planner-level guarantees.

1. Terminological scope and conceptual variants

The literature uses DOD in more than one operational sense. In image generation and image restoration, the most literal meaning is single-step sampling: a model maps a noisy latent directly to a clean latent or restored image in one forward pass (xu et al., 2023, Tang et al., 5 Aug 2025). In anomaly detection, the same once-and-done principle is realized as single-step scoring rather than single-step generation: OSD-IRF computes an inverse residual field from a single call to the DDPM noise predictor and evaluates Gaussian typicality in that space (Zhang et al., 20 Apr 2026). In kinodynamic motion planning, the term is not used explicitly, but the paper states that DiTree embodies the intended sense of DOD because a diffusion-based policy is trained once in a single environment and then reused across unseen environments and robot systems without retraining, while the sampling-based planner preserves safety and probabilistic completeness (Hassidof et al., 28 Aug 2025).

A common misconception is that DOD is synonymous with distillation. The cited works show otherwise. The all-in-one restoration paper adopts distribution matching distillation to enable one-step restoration (Tang et al., 5 Aug 2025), but UFOGen uses a diffusion-GAN hybrid objective rather than ODE solvers or distillation (xu et al., 2023), YOSO trains a one-step denoiser with self-cooperative adversarial learning (Luo et al., 2024), OSD-IRF avoids distillation and exploits an analytical residual-field construction (Zhang et al., 20 Apr 2026), and "On the Design of One-step Diffusion via Shortcutting Flow Paths" reports a one-step model with FID50k of 2.85 on ImageNet-256x256 that requires no pre-training, distillation, or curriculum learning (Lin et al., 3 Dec 2025).

A second misconception is that DOD always concerns image synthesis. The surveyed papers place the paradigm in at least four domains: text-to-image generation (xu et al., 2023, Luo et al., 2024), all-in-one image restoration (Tang et al., 5 Aug 2025), unsupervised industrial anomaly detection (Zhang et al., 20 Apr 2026), and kinodynamic motion planning (Hassidof et al., 28 Aug 2025). This suggests that DOD is better understood as an inference-time principle than as a domain-specific model class.

2. Mathematical mechanisms for one-step diffusion

One family of DOD methods modifies diffusion training so that the learned model predicts a clean sample directly. UFOGen operates in the latent space of Stable Diffusion 1.5 and changes both the generator parameterization and the reconstruction target. Its objective applies an adversarial loss to the noisy marginal at $x_{t-1}$ while explicitly reconstructing clean data at $x_0$:

$$\begin{aligned} &\min_\theta\max_{D_\phi}\mathbb{E}_{q(x_0)q(x_{t-1}|x_0), p_{\theta}(x_{0}')p_{\theta}(x_{t-1}'|x_0')} \Bigl[ \log(D_{\phi}(x_{t-1}, t)) + \log(1-D_{\phi}(x'_{t-1}, t)) \ &\quad + \lambda_{KL}\gamma_t\left\lVert x_0-x_0'\right\rVert^2 \Bigl]. \end{aligned}$$

The paper argues that adversarial matching at $x_{t-1}$ implicitly matches the clean distributions because both marginals are convolutions with the same Gaussian kernel, while the direct $\|x_0-x_0'\|^2$ term makes clean alignment explicit and lower-variance (xu et al., 2023).

YOSO also uses a diffusion-GAN hybrid, but it relocates adversarial matching to clean marginals $p_\theta^{(t)}(x_0)$ and smooths the target distribution by using the model’s own less-corrupted clean marginal $p_\theta^{(t-1)}(x_0)$ as the discriminator’s reference. This self-cooperative construction is designed to reduce discriminator overpowering and mode collapse relative to matching directly against $q(x_0)$, and the final objective adds both pointwise denoising and consistency regularization (Luo et al., 2024). In this formulation, DOD emerges from a one-shot denoising generator trained under a smoothed adversarial divergence on clean marginals rather than from iterative reverse diffusion.

A second family of methods reaches one-step diffusion through shortcutting probability flow paths. The shortcut-model framework defines flow maps $X_{t,r}$ that jump from time $t$ to time $x_0$0 directly and trains them by enforcing the consistency relation

$x_0$1

The paper provides Wasserstein-$x_0$2 bounds of the form

$x_0$3

thereby giving a theoretical justification for one-step generation via shortcut objectives (Lin et al., 3 Dec 2025). The continuous-time shortcut setting removes the explicit $x_0$4 term and is reported to yield tighter control and better empirical fidelity.

A third mechanism is task-specific distillation. The restoration paper adopts DMD to distill multi-step latent denoising into a single reverse update and then recovers the clean latent by

$x_0$5

This one-step estimate is then decoded by a VAE decoder augmented with detail enhancement modules (Tang et al., 5 Aug 2025).

A fourth mechanism dispenses with generation entirely and performs one-step scoring. OSD-IRF defines the normalized residual field

$x_0$6

so that $x_0$7 for any neighboring time step. The inverse residual field is estimated by the DDPM noise predictor, either as $x_0$8 or, in the paper’s mean-path variant, as

$x_0$9

Anomaly scoring then reduces to Gaussian likelihood evaluation in IRF space (Zhang et al., 20 Apr 2026).

3. One-step image synthesis

In large-scale text-to-image generation, UFOGen instantiates DOD as "You Forward Once." It uses a UNet generator and UNet discriminator in Stable Diffusion 1.5 latent space, with the frozen CLIP ViT-L/14 text encoder for conditioning and the SD VAE for decoding. Both generator and discriminator are initialized from the pre-trained SD 1.5 UNet. Inference is a single forward pass: sample $$\begin{aligned} &\min_\theta\max_{D_\phi}\mathbb{E}_{q(x_0)q(x_{t-1}|x_0), p_{\theta}(x_{0}')p_{\theta}(x_{t-1}'|x_0')} \Bigl[ \log(D_{\phi}(x_{t-1}, t)) + \log(1-D_{\phi}(x'_{t-1}, t)) \ &\quad + \lambda_{KL}\gamma_t\left\lVert x_0-x_0'\right\rVert^2 \Bigl]. \end{aligned}$$0, compute $$\begin{aligned} &\min_\theta\max_{D_\phi}\mathbb{E}_{q(x_0)q(x_{t-1}|x_0), p_{\theta}(x_{0}')p_{\theta}(x_{t-1}'|x_0')} \Bigl[ \log(D_{\phi}(x_{t-1}, t)) + \log(1-D_{\phi}(x'_{t-1}, t)) \ &\quad + \lambda_{KL}\gamma_t\left\lVert x_0-x_0'\right\rVert^2 \Bigl]. \end{aligned}$$1, and decode with the SD VAE (xu et al., 2023). On MSCOCO-2017 5k zero-shot, UFOGen (1 step) reports time 0.09 s, FID-5k 22.5, and CLIP 0.311; on MSCOCO-2014 30k, it reports time 0.09 s, 0.9B parameters, and FID-30k 12.78 (xu et al., 2023).

YOSO frames the same objective as "You Only Sample Once." For text-to-image, it initializes the generator from PixArt-$$\begin{aligned} &\min_\theta\max_{D_\phi}\mathbb{E}_{q(x_0)q(x_{t-1}|x_0), p_{\theta}(x_{0}')p_{\theta}(x_{t-1}'|x_0')} \Bigl[ \log(D_{\phi}(x_{t-1}, t)) + \log(1-D_{\phi}(x'_{t-1}, t)) \ &\quad + \lambda_{KL}\gamma_t\left\lVert x_0-x_0'\right\rVert^2 \Bigl]. \end{aligned}$$2 or Stable Diffusion 1.5, uses a latent discriminator built from the pre-trained Stable Diffusion 1.5 UNet, and introduces a latent perceptual loss computed with SD HalfUNet features at $$\begin{aligned} &\min_\theta\max_{D_\phi}\mathbb{E}_{q(x_0)q(x_{t-1}|x_0), p_{\theta}(x_{0}')p_{\theta}(x_{t-1}'|x_0')} \Bigl[ \log(D_{\phi}(x_{t-1}, t)) + \log(1-D_{\phi}(x'_{t-1}, t)) \ &\quad + \lambda_{KL}\gamma_t\left\lVert x_0-x_0'\right\rVert^2 \Bigl]. \end{aligned}$$3. It also addresses flawed terminal schedulers through Informative Prior Initialization and a quick adaptation stage that switches to $$\begin{aligned} &\min_\theta\max_{D_\phi}\mathbb{E}_{q(x_0)q(x_{t-1}|x_0), p_{\theta}(x_{0}')p_{\theta}(x_{t-1}'|x_0')} \Bigl[ \log(D_{\phi}(x_{t-1}, t)) + \log(1-D_{\phi}(x'_{t-1}, t)) \ &\quad + \lambda_{KL}\gamma_t\left\lVert x_0-x_0'\right\rVert^2 \Bigl]. \end{aligned}$$4-prediction and zero terminal SNR (Luo et al., 2024). On HPS v2.1 and AeS, YOSO-PixArt-$$\begin{aligned} &\min_\theta\max_{D_\phi}\mathbb{E}_{q(x_0)q(x_{t-1}|x_0), p_{\theta}(x_{0}')p_{\theta}(x_{t-1}'|x_0')} \Bigl[ \log(D_{\phi}(x_{t-1}, t)) + \log(1-D_{\phi}(x'_{t-1}, t)) \ &\quad + \lambda_{KL}\gamma_t\left\lVert x_0-x_0'\right\rVert^2 \Bigl]. \end{aligned}$$5 at one step reports HPS 28.21 and AeS 6.18; YOSO-SD 1.5 reports 28.11 and 6.25. With LoRA fine-tuning, YOSO-LoRA at one step reports HPS 23.34 and AeS 5.73 (Luo et al., 2024).

The shortcutting-flow-path literature places one-step image synthesis on a different footing: one-step generation is learned from scratch rather than obtained by finetuning a pre-trained diffusion backbone or by introducing an adversarial discriminator. The paper’s ESC configuration, built from continuous-time shortcutting, linear paths, variance reduction via a plug-in velocity, a gradual time sampler, and adaptive loss weighting, achieves 1-NFE FID50k = 2.85 on ImageNet-256x256 under classifier-free guidance, with no pre-training, distillation, or curriculum learning (Lin et al., 3 Dec 2025). This result is significant because earlier one-step-from-scratch baselines in the same comparison are substantially weaker, including CT at FID50k $$\begin{aligned} &\min_\theta\max_{D_\phi}\mathbb{E}_{q(x_0)q(x_{t-1}|x_0), p_{\theta}(x_{0}')p_{\theta}(x_{t-1}'|x_0')} \Bigl[ \log(D_{\phi}(x_{t-1}, t)) + \log(1-D_{\phi}(x'_{t-1}, t)) \ &\quad + \lambda_{KL}\gamma_t\left\lVert x_0-x_0'\right\rVert^2 \Bigl]. \end{aligned}$$6, IMM at $$\begin{aligned} &\min_\theta\max_{D_\phi}\mathbb{E}_{q(x_0)q(x_{t-1}|x_0), p_{\theta}(x_{0}')p_{\theta}(x_{t-1}'|x_0')} \Bigl[ \log(D_{\phi}(x_{t-1}, t)) + \log(1-D_{\phi}(x'_{t-1}, t)) \ &\quad + \lambda_{KL}\gamma_t\left\lVert x_0-x_0'\right\rVert^2 \Bigl]. \end{aligned}$$7, and SCD at $$\begin{aligned} &\min_\theta\max_{D_\phi}\mathbb{E}_{q(x_0)q(x_{t-1}|x_0), p_{\theta}(x_{0}')p_{\theta}(x_{t-1}'|x_0')} \Bigl[ \log(D_{\phi}(x_{t-1}, t)) + \log(1-D_{\phi}(x'_{t-1}, t)) \ &\quad + \lambda_{KL}\gamma_t\left\lVert x_0-x_0'\right\rVert^2 \Bigl]. \end{aligned}$$8 (Lin et al., 3 Dec 2025).

Taken together, these papers delineate three distinct synthesis routes to DOD: adversarial clean-latent matching with explicit reconstruction (xu et al., 2023), self-cooperative adversarial learning over clean marginals (Luo et al., 2024), and shortcut consistency training over probability flow maps (Lin et al., 3 Dec 2025).

4. All-in-one image restoration

The paper that uses the name DOD explicitly defines it as an efficient all-in-one image restoration method that achieves superior restoration performance with only one-step sampling of Stable Diffusion models (Tang et al., 5 Aug 2025). Its target setting is AiOIR, where a single model must handle heterogeneous degradations such as noise, rain, haze, blur, and low-light without relying on textual prompts.

The architecture retains the latent diffusion backbone of Stable Diffusion but adds three task-specific components. Multi-degradation Feature Modulation extracts degradation condition vectors $$\begin{aligned} &\min_\theta\max_{D_\phi}\mathbb{E}_{q(x_0)q(x_{t-1}|x_0), p_{\theta}(x_{0}')p_{\theta}(x_{t-1}'|x_0')} \Bigl[ \log(D_{\phi}(x_{t-1}, t)) + \log(1-D_{\phi}(x'_{t-1}, t)) \ &\quad + \lambda_{KL}\gamma_t\left\lVert x_0-x_0'\right\rVert^2 \Bigl]. \end{aligned}$$9 from the intermediate $x_{t-1}$0-space of a pretrained DDPM at timestep $x_{t-1}$1, motivated by the empirical finding that these features are highly discriminative for different degradations (Tang et al., 5 Aug 2025). Parameter-efficient Conditional LoRA then modulates the low-rank pathway by degradation-aware affine parameters:

$x_{t-1}$2

with per-layer parameters produced from $x_{t-1}$3 by a small mapper $x_{t-1}$4 (Tang et al., 5 Aug 2025). Finally, High-fidelity Detail Enhancement inserts RRDB-based Detail Enhancement Modules into the VAE decoder:

$x_{t-1}$5

Training is split into two stages. Stage 1 trains PLA/CLoRA while freezing the VAE decoder, combining a data term with DMD-based distribution matching. Stage 2 freezes PLA/CLoRA and trains only the decoder-side DEMs with reconstruction and SSIM losses (Tang et al., 5 Aug 2025). Ablation identifies UNet-only conditional placement as preferable, and LoRA rank $x_{t-1}$6 as the best trade-off; $x_{t-1}$7 collapses training, while $x_{t-1}$8 degrades slightly (Tang et al., 5 Aug 2025).

Empirically, the paper reports a three-task average over dehazing, deraining, and denoising at $x_{t-1}$9 of PSNR 31.87, SSIM 0.9122, LPIPS 0.0620, DISTS 0.0626, CLIPIQA 0.6660, NIQE 3.5212, MUSIQ 68.85, and MANIQA 0.7080 (Tang et al., 5 Aug 2025). On five-task perceptual evaluation, it reports MUSIQ/MANIQA values of 67.99 / 0.7115 for dehazing, 70.49 / 0.7104 for deraining, 69.50 / 0.7219 for denoising at $\|x_0-x_0'\|^2$0, 37.77 / 0.4286 for deblurring, and 73.75 / 0.6657 for low-light (Tang et al., 5 Aug 2025). At 512×512 on RTX 3090, DOD reports 0.211 s inference time versus 17.664 s for DA-CLIP, 26.602 s for AutoDIR, 212.534 s for WeatherDiff, and 0.303 s for DiffUIR (Tang et al., 5 Aug 2025). The paper summarizes this as approximately 84× faster than DA-CLIP while delivering superior perceptual quality (Tang et al., 5 Aug 2025).

The restoration setting also clarifies a central feature of DOD in low-level vision: one-step inference alone is usually insufficient. The method compensates by combining degradation-aware conditioning, parameter-efficient UNet adaptation, and decoder-side detail recovery (Tang et al., 5 Aug 2025).

5. Beyond image synthesis: anomaly detection and motion planning

OSD-IRF extends DOD to unsupervised industrial anomaly detection. The method trains an unconditional DDPM on normal data only, extracts features with EfficientNet-B4, and then evaluates the DDPM noise predictor once at a fixed time step to obtain the inverse residual field. Anomaly detection is performed by evaluating the probability density of the IRF under a Gaussian distribution and thresholding the score; with $\|x_0-x_0'\|^2$1, the image-level score reduces to $\|x_0-x_0'\|^2$2 plus a constant (Zhang et al., 20 Apr 2026). The paper reports that this requires only single step diffusion because the IRF holds for any neighboring time step in the denoising process. On MVTec-AD it reports image AUROC 99.0, AP 99.6, F1-max 98.4; pixel AUROC 97.7, AP 54.3, F1-max 57.4, AU-PRO 93.5; mAD 85.7; FPS 133. On VisA it reports mAD 81.3 and FPS 105.9, and on MPDD mAD 83.8 and FPS 212.4 (Zhang et al., 20 Apr 2026). The text summarizes the speed gain as roughly a 2X inference speedup without distillation (Zhang et al., 20 Apr 2026).

DiTree transfers the once-and-done intuition to kinodynamic planning. The method trains a diffusion policy once in a single environment, the D4RL AntMaze Large map, and then reuses it across unseen maps and two robot systems without retraining (Hassidof et al., 28 Aug 2025). The policy samples short action sequences

$\|x_0-x_0'\|^2$3

conditioned on the current tree state, a target encoded relative to that state, and a local occupancy observation; an RRT backbone then forward-propagates the sampled controls under the system dynamics, collision-checks the resulting segment, and inserts only feasible edges into the search tree (Hassidof et al., 28 Aug 2025). The paper emphasizes that safety is enforced by collision checking and that returned solutions are collision-free by construction. It also states Theorem 1: for a Lipschitz-continuous system and a full-support action sampler, RRT-based DiTree is probabilistically complete, with solution probability at least $\|x_0-x_0'\|^2$4 after $\|x_0-x_0'\|^2$5 samples for some constants $\|x_0-x_0'\|^2$6 (Hassidof et al., 28 Aug 2025).

The empirical evaluation trains once in a single environment and tests on 15 distinct unseen maps with 20 trials per scenario and a 120 s time limit per query. Aggregate CarMaze results are 73.3% success and 23.3 s runtime for DiTree, 45.0% and 67.3 s for RRT, 47.3% and 66.2 s for SST, and 46.0% and 20.7 s for DP-only. Aggregate AntMaze results are 35.7% success and 59.4 s runtime for DiTree, 7.3% and 54.5 s for DP-only, while classical SBPs fail entirely in that setup (Hassidof et al., 28 Aug 2025). The paper summarizes DiTree as about 3× faster than classical SBPs on average and roughly 30% higher in success rate than all baselines across tasks (Hassidof et al., 28 Aug 2025).

These two papers broaden the meaning of DOD. In OSD-IRF, diffusion is executed once to produce a tractable diagnostic latent. In DiTree, the diffusion model is trained once and then repeatedly reused as an informed sampler inside a planner whose guarantees are inherited from the sampling-based planning backbone.

6. Guarantees, limitations, and open questions

The strongest formal guarantees in this literature appear in the shortcut-model and planning variants. Shortcutting-flow-path methods provide explicit Wasserstein-$\|x_0-x_0'\|^2$7 control for the learned one-step generator under discrete-time and continuous-time shortcut losses (Lin et al., 3 Dec 2025). DiTree provides planner-level safety through collision checking, kinodynamic feasibility through forward propagation under $\|x_0-x_0'\|^2$8, and probabilistic completeness under Lipschitz dynamics and full-support action sampling; the paper further states that asymptotic optimality can be inherited by swapping the backbone to AO-RRT or SST (Hassidof et al., 28 Aug 2025). By contrast, UFOGen and YOSO focus on training objectives and empirical quality rather than formal sampling guarantees (xu et al., 2023, Luo et al., 2024).

Training stability is a recurring issue. UFOGen introduces adversarial matching on a noisy marginal and explicit clean reconstruction because conventional one-step diffusion often sacrifices fidelity or requires multiple steps (xu et al., 2023). YOSO argues that directly matching generated clean marginals against real clean data destabilizes the discriminator and harms one-step efficacy; its ablations report FID 10.85 for direct matching to real data and 45.15 for matching over corrupted data, versus 3.82 for the self-cooperative design on CIFAR-10 (Luo et al., 2024). The restoration paper addresses a different stability problem: one-step sampling tends to lose fine textures, so it supplements the denoiser with RRDB-based decoder enhancements (Tang et al., 5 Aug 2025).

The limitations are domain-specific. UFOGen notes missing objects, attribute leakage, and counting errors similar to SD-based models, and does not describe safety or content filtering measures (xu et al., 2023). YOSO identifies scheduler flaws as a source of artifacts in one-step generation and introduces Informative Prior Initialization plus quick adaptation to zero terminal SNR as remedies (Luo et al., 2024). The restoration paper states that single-step sampling can under-restore extremely subtle textures or very challenging compound degradations without DEM, and that unusual artifacts not represented in the pretrained DDPM may weaken degradation conditioning (Tang et al., 5 Aug 2025). OSD-IRF assumes anomalies are off-manifold and depends on accurate $\|x_0-x_0'\|^2$9; distribution shift in the normal-data training regime can degrade IRF typicality (Zhang et al., 20 Apr 2026). DiTree assumes full, static obstacle maps, can slow under severe distribution shifts or pathological local traps, and shifts the computational bottleneck from collision checking to diffusion inference (Hassidof et al., 28 Aug 2025).

A final conceptual clarification concerns what is being made "once and done." In synthesis papers, it is the reverse diffusion trajectory itself (xu et al., 2023, Luo et al., 2024, Lin et al., 3 Dec 2025). In restoration, it is the reverse update from degraded latent to restored latent (Tang et al., 5 Aug 2025). In anomaly detection, it is the computation of a score-bearing latent variable rather than a reconstructed sample (Zhang et al., 20 Apr 2026). In planning, it is the training phase of the diffusion prior, not the planner’s search process, which still unfolds as an SBP tree expansion (Hassidof et al., 28 Aug 2025). This suggests that DOD is not a single algorithmic recipe but a broader compression principle: the diffusion component is designed so that one evaluation, or one training investment followed by universal reuse, suffices for deployment.