ZADS: Zero-shot Adaptive Diffusion Sampling

• ZADS is a methodological pattern that uses a fixed, pre-trained diffusion prior with adaptive test-time modifications based on measurements and uncertainty.
• It enables adaptive control in inverse problems such as MRI, segmentation, and compressed sensing by adjusting sampling trajectories without retraining the generative model.
• Different variants of ZADS modify the reverse process through adaptive measurement selection, timestep tuning, or analytic corrections, enhancing reconstruction performance across domains.

Zero-shot Adaptive Diffusion Sampling (ZADS) denotes a family of diffusion-based inference procedures in which a pre-trained generative prior is reused without retraining, while the sampling trajectory is adapted at inference time using measurements, physics constraints, posterior uncertainty, or timestep-dependent controls. The term is not fully standardized. In AdaSense, “Zero-shot Adaptive Diffusion Sampling (ZADS)” is a conceptual name for adaptive compressed sensing via posterior sampling (Elata et al., 2024). In later MRI work, ZADS becomes the official acronym for test-time optimization of per-timestep fidelity weights (Alçalar et al., 11 Sep 2025). Closely related papers use neighboring labels such as ZAPS and AdaDS for zero-shot adaptation of posterior samplers, adaptive timestep selection, or uncertainty-conditioned denoising (Alçalar et al., 2024, Wang et al., 10 Feb 2026, Nakhli et al., 14 Jun 2026). This suggests that ZADS is best understood as a methodological pattern rather than a single canonical algorithm.

1. Nomenclature and conceptual scope

In the AdaSense formulation, ZADS is defined as the core idea of using “a pre-trained diffusion model zero-shot as a Bayesian prior, draw posterior samples conditioned on currently acquired measurements, and adaptively choose the next measurements to maximally reduce uncertainty” (Elata et al., 2024). That definition places ZADS at the intersection of Bayesian experimental design, compressed sensing, and diffusion-based posterior sampling.

Subsequent work broadens the meaning. In ZAPS, the paper explicitly states that “the requested term ZADS is not the official acronym in the paper but conceptually aligns with ZAPS,” because ZAPS performs “zero-shot, per-timestep adaptive weighting of the likelihood guidance and a curvature-aware adaptation” for inverse problems (Alçalar et al., 2024). In AdaDS for depth super-resolution, the paper likewise states that “the term ‘Zero-shot Adaptive Diffusion Sampling (ZADS)’ is not explicitly used,” but that AdaDS “is the instantiation of this concept” through adaptive timestep selection and tailored noise injection (Wang et al., 10 Feb 2026). In zero-shot segmentation, ZADS is described as a training-free procedure that “adaptively select[s], per pixel, the denoising timestep that best captures the desired semantic granularity” (Nakhli et al., 14 Jun 2026).

The strongest terminological standardization occurs in the MRI paper titled “Automated Tuning for Diffusion Inverse Problem Solvers without Generative Prior Retraining,” where ZADS is the official method name. There it refers to “a test-time optimization framework that turns a diffusion-based inverse solver into a fixed unrolled sampler and then automatically tunes per-timestep data fidelity weights using only the undersampled measurements” (Alçalar et al., 11 Sep 2025).

Across these usages, the common elements are stable: a frozen diffusion prior, no retraining of that prior, and some adaptive mechanism defined entirely at inference time. What changes across papers is the adaptation target: measurement selection, fidelity weights, starting timestep, structured noise, or semantic feature extraction.

2. Shared probabilistic structure

Most ZADS-style methods begin from a known forward model, typically linear. AdaSense writes the compressed sensing observation model as

$$y = A x + \epsilon,\qquad \epsilon \sim \mathcal{N}(0,\sigma^2 I),$$

with $x \in \mathbb{R}^D$, $y \in \mathbb{R}^d$, and $d < D$ (Elata et al., 2024). Under a Gaussian likelihood,

$p(x \mid y) \propto p(y \mid x)\,p(x),$

and likelihood-guided sampling augments the prior score with a data term. In the generic score-based formulation used in AdaSense,

$$\nabla_x \log p(x \mid y) = s_\theta(x,t) + \frac{1}{\sigma^2} A^\top (y - A x),$$

where $s_\theta(x,t) \approx \nabla_x \log p_t(x)$ is supplied by a pre-trained diffusion model (Elata et al., 2024).

Related zero-shot samplers modify this template differently. ZAPS fixes a small number of reverse steps and learns per-timestep log-likelihood weights $\{\zeta_t\}$ together with a wavelet-diagonal approximation of the prior Hessian, represented as $\mathbf{W}\mathbf{D}_t\mathbf{W}^\top$, using only a physics-guided loss $\|\mathbf{A}\mathbf{x}_0-\mathbf{y}\|_2^2$ at inference time (Alçalar et al., 2024). The MRI ZADS formulation keeps the pretrained score model fixed but makes the data-consistency strength timestep-specific, with a separate scalar $x \in \mathbb{R}^D$0 for each chosen timestep in an unrolled DDS sampler (Alçalar et al., 11 Sep 2025). AdaDS for depth does not add an explicit likelihood score during denoising; instead it chooses an uncertainty-dependent reverse starting timestep and injects tailored noise before a one-step LCM denoising pass (Wang et al., 10 Feb 2026).

A theoretical account of this family appears in work on score-mismatched zero-shot conditional samplers. There, score mismatch produces “an asymptotic distributional bias between the target and sampling distributions, proportional to the accumulated mismatch between the target and training distributions,” and the dominant term takes the form

$x \in \mathbb{R}^D$1

where $x \in \mathbb{R}^D$2 denotes the per-step mismatch between the target reverse mean and the sampler’s reverse mean (Liang et al., 2024). In that perspective, ZADS mechanisms can be interpreted as attempts to reduce mismatch by adapting the sampler to the measurement condition, schedule, or uncertainty profile.

3. Adaptive acquisition in compressed sensing: AdaSense

AdaSense is the clearest instance of ZADS as adaptive measurement design. It seeks to reconstruct an unknown image or signal from a budgeted sequence of linear measurements, while choosing future measurements sequentially using posterior uncertainty estimated from diffusion posterior samples. The method is training-free for the inverse operator and uses a pre-trained diffusion prior together with zero-shot posterior sampling, instantiated in the paper with DDRM (Elata et al., 2024).

For the noiseless case and the paper’s parameterization, AdaSense uses the single-step conditional DDRM update

$x \in \mathbb{R}^D$3

which enforces measurement consistency at each step. At adaptive stage $x \in \mathbb{R}^D$4, the method draws $x \in \mathbb{R}^D$5 posterior samples from $x \in \mathbb{R}^D$6, centers them, estimates empirical posterior uncertainty, and chooses the next $x \in \mathbb{R}^D$7 measurements. In the unconstrained case, the next directions are the top $x \in \mathbb{R}^D$8 eigenvectors of the empirical posterior covariance. In the constrained case, where measurements must come from a finite candidate set $x \in \mathbb{R}^D$9 such as MRI k-space lines or CT angles, AdaSense uses a sample-based surrogate that selects candidates whose projections capture the most posterior energy.

The acquisition criterion is posterior variance maximization. For a candidate probe $y \in \mathbb{R}^d$0,

$y \in \mathbb{R}^d$1

and under a Gaussian approximation this becomes $y \in \mathbb{R}^d$2. The paper notes that maximizing posterior variance along $y \in \mathbb{R}^d$3 also maximizes mutual information when the measurement noise is Gaussian.

Domain	Setting	Reported outcome
Faces	CelebA-HQ 256×256, 192 measurements	AdaSense PSNR 26.2, SSIM 0.7515, LPIPS 0.1950; PCA PSNR 24.6
Faces	Mean of 16 posterior samples	PSNR 26.87, SSIM 0.7713, LPIPS 0.2449
MRI	Vertical subsampling, R10	PSNR 27.01, SSIM 0.5229
MRI	General subsampling, R400	PSNR 25.26, SSIM 0.4124 vs Poisson disk 23.06, 0.3728
MRI	General subsampling, R200	PSNR 25.94, SSIM 0.4331 vs Poisson disk 24.31, 0.3999
MRI	Active acquisition, R8–30L	PSNR 28.89, SSIM 0.6108
MRI	Active acquisition, R16–2L	PSNR 27.51, SSIM 0.5547
CT	DeepLesion sparse-view CT	Adaptive angle selection improves PSNR/SSIM over non-adaptive angle selections

The paper also reports that, for a fixed budget $y \in \mathbb{R}^d$4, increasing $y \in \mathbb{R}^d$5 improves PSNR and LPIPS, indicating that more frequent adaptation is beneficial. Computationally, AdaSense uses roughly $y \in \mathbb{R}^d$6 NFEs per posterior sample with DDRM, and the posterior sample count ranges from approximately $y \in \mathbb{R}^d$7 to $y \in \mathbb{R}^d$8 depending on the domain and acceleration. The stated heuristic is $y \in \mathbb{R}^d$9 for estimating $d < D$0.

AdaSense therefore exemplifies one historically important meaning of ZADS: adaptive experimental design driven directly by diffusion-based posterior uncertainty, without a learned acquisition policy and without retraining across modalities such as faces, MRI, and CT.

4. Per-timestep adaptation and self-supervised sampler tuning

A second line of work uses ZADS not to choose measurements but to tune the reverse process itself. ZAPS is representative: it keeps a fixed irregular timestep schedule, then performs zero-shot training at inference time to learn per-timestep likelihood weights $d < D$1 and diagonal curvature surrogates $d < D$2 under a physics-guided loss. The reverse update adds an adaptive guidance term of the form

$d < D$3

with no backpropagation through the frozen score network (Alçalar et al., 2024).

ZAPS targets the instability of heuristic likelihood weights under irregular schedules. The reported typical setting is $d < D$4 steps and $d < D$5 epochs, for about $d < D$6 NFEs instead of the $d < D$7 steps often used by DPS. The paper reports wall-clock time on a single A100 GPU of approximately $d < D$8 s for ZAPS versus $d < D$9 s for DPS, together with improved or competitive LPIPS, SSIM, and PSNR across Gaussian deblurring, motion deblurring, inpainting, and super-resolution. On FFHQ Gaussian deblurring, for example, ZAPS reports LPIPS $p(x \mid y) \propto p(y \mid x)\,p(x),$0, SSIM $p(x \mid y) \propto p(y \mid x)\,p(x),$1, and PSNR $p(x \mid y) \propto p(y \mid x)\,p(x),$2, versus DPS at $p(x \mid y) \propto p(y \mid x)\,p(x),$3, $p(x \mid y) \propto p(y \mid x)\,p(x),$4, and $p(x \mid y) \propto p(y \mid x)\,p(x),$5.

The official ZADS method in accelerated MRI applies the same broad principle to DDS-style sampling. It replaces the single global fidelity parameter with timestep-specific weights $p(x \mid y) \propto p(y \mid x)\,p(x),$6 and optimizes them at test time using only undersampled k-space data. To avoid overfitting, the sampled k-space set $p(x \mid y) \propto p(y \mid x)\,p(x),$7 is split SSDU-style into $p(x \mid y) \propto p(y \mid x)\,p(x),$8 for data consistency and $p(x \mid y) \propto p(y \mid x)\,p(x),$9 for held-out supervision. At timestep $$\nabla_x \log p(x \mid y) = s_\theta(x,t) + \frac{1}{\sigma^2} A^\top (y - A x),$$0, the reconstruction is refined by a CG solve

$$\nabla_x \log p(x \mid y) = s_\theta(x,t) + \frac{1}{\sigma^2} A^\top (y - A x),$$1

and the learned weights are optimized using the normalized held-out loss

$$\nabla_x \log p(x \mid y) = s_\theta(x,t) + \frac{1}{\sigma^2} A^\top (y - A x),$$2

On fastMRI multi-coil knee reconstruction with equidistant $$\nabla_x \log p(x \mid y) = s_\theta(x,t) + \frac{1}{\sigma^2} A^\top (y - A x),$$3 undersampling, the paper reports PD performance of $$\nabla_x \log p(x \mid y) = s_\theta(x,t) + \frac{1}{\sigma^2} A^\top (y - A x),$$4 dB and SSIM $$\nabla_x \log p(x \mid y) = s_\theta(x,t) + \frac{1}{\sigma^2} A^\top (y - A x),$$5, and PD-FS performance of $$\nabla_x \log p(x \mid y) = s_\theta(x,t) + \frac{1}{\sigma^2} A^\top (y - A x),$$6 dB and SSIM $$\nabla_x \log p(x \mid y) = s_\theta(x,t) + \frac{1}{\sigma^2} A^\top (y - A x),$$7, outperforming $$\nabla_x \log p(x \mid y) = s_\theta(x,t) + \frac{1}{\sigma^2} A^\top (y - A x),$$8-wavelet CS, DPS, and DDS under the reported protocol (Alçalar et al., 11 Sep 2025).

A third variant replaces weight tuning by analytic guidance. “Zero-Shot Solving of Imaging Inverse Problems via Noise-Refined Likelihood Guided Diffusion Models” derives a closed-form approximation to the likelihood score and uses it to refine the model-predicted noise before a DDIM step. The refined noise is

$$\nabla_x \log p(x \mid y) = s_\theta(x,t) + \frac{1}{\sigma^2} A^\top (y - A x),$$9

where the approximate likelihood score is evaluated around the Tweedie estimate $s_\theta(x,t) \approx \nabla_x \log p_t(x)$0 (Wang et al., 16 Jun 2025). In compressive sensing at a $s_\theta(x,t) \approx \nabla_x \log p_t(x)$1 sampling rate, the paper reports on CelebA-HQ a DD-NRLG PSNR of $s_\theta(x,t) \approx \nabla_x \log p_t(x)$2 dB and SSIM $s_\theta(x,t) \approx \nabla_x \log p_t(x)$3, and an ID-NRLG PSNR of $s_\theta(x,t) \approx \nabla_x \log p_t(x)$4 dB, SSIM $s_\theta(x,t) \approx \nabla_x \log p_t(x)$5, and LPIPS $s_\theta(x,t) \approx \nabla_x \log p_t(x)$6. The same paper reports DD-NRLG runtime of about $s_\theta(x,t) \approx \nabla_x \log p_t(x)$7 seconds per image for 100 steps, compared with DPS at about $s_\theta(x,t) \approx \nabla_x \log p_t(x)$8 seconds for 1000 steps.

Taken together, these methods define another major meaning of ZADS: adaptive, zero-shot control of the reverse diffusion dynamics, either by learned timestep-wise guidance parameters or by closed-form likelihood-driven corrections.

5. Adaptive timesteps and cross-domain generalizations

ZADS-style ideas have also moved beyond classical inverse problems into depth estimation, segmentation, and protein design. In these settings, the adaptive variable is often not a measurement direction or a fidelity scalar, but the denoising timestep itself or an auxiliary sampling correction.

AdaDS formulates robust depth super-resolution under arbitrary or unknown degradation. The method uses a calibration network $s_\theta(x,t) \approx \nabla_x \log p_t(x)$9 to produce a refined latent $\{\zeta_t\}$0 and per-pixel uncertainty $\{\zeta_t\}$1, then chooses a reverse starting timestep through the rule

$\{\zeta_t\}$2

clamped to $\{\zeta_t\}$3, with $\{\zeta_t\}$4 in the reported experiments (Wang et al., 10 Feb 2026). It then injects tailored noise and performs one-step denoising with Marigold-LCM. The paper attributes the procedure to the contraction property of Gaussian smoothing and reports that AdaDS improves RMSE by $\{\zeta_t\}$5 over the second-best PriorDA on RGB-D-D at native LR, surpasses C2PD by $\{\zeta_t\}$6 RMSE on TOFDSR, and achieves average gains over PromptDA of $\{\zeta_t\}$7 RMSE on ScanNet and $\{\zeta_t\}$8 RMSE on NYUv2.

In zero-shot semantic segmentation, adaptive timestep selection becomes spatially local. The segmentation paper constructs Contextual Similarity Maps (CSMs) by fusing multi-resolution self-attention with deep U-Net encoder features,

$\{\zeta_t\}$9

then uses a Temporal Stability Matrix, a Fiedler-vector-based spectral analysis, and PELT change-point detection to choose a per-pixel target timestep $\mathbf{W}\mathbf{D}_t\mathbf{W}^\top$0 (Nakhli et al., 14 Jun 2026). The paper argues that semantics evolve from part-level abstractions at earlier timesteps to object-level abstractions at later stages, and that static timestep selection therefore underuses the diffusion hierarchy. With a 22-timestep setting on SDv1.4, the method reports mIoU values of $\mathbf{W}\mathbf{D}_t\mathbf{W}^\top$1 on VOC, $\mathbf{W}\mathbf{D}_t\mathbf{W}^\top$2 on Context, $\mathbf{W}\mathbf{D}_t\mathbf{W}^\top$3 on ADE20K, $\mathbf{W}\mathbf{D}_t\mathbf{W}^\top$4 on COCO-Stuff, and $\mathbf{W}\mathbf{D}_t\mathbf{W}^\top$5 on Cityscapes, all under the reported zero-shot protocol.

Protein motif scaffolding offers a different extension. The paper on SMC-aided diffusion posterior sampling treats an unconditional protein diffusion model, Genie, as a prior and composes motif constraints through likelihood or energy potentials. Reconstruction guidance is then corrected by Sequential Monte Carlo, with ESS-triggered resampling when $\mathbf{W}\mathbf{D}_t\mathbf{W}^\top$6 (Young et al., 2024). The paper states that this case study exemplifies zero-shot, adaptively corrected diffusion posterior sampling. In the single-motif benchmark, $\mathbf{W}\mathbf{D}_t\mathbf{W}^\top$7 tasks had at least one successful solution among 32 samples, and TDS-rmsd solved $\mathbf{W}\mathbf{D}_t\mathbf{W}^\top$8 problems. In selected multi-motif tasks with 512 samples, TDS-rmsd found one unique success in each of two tasks. Here ZADS is instantiated as zero-shot posterior steering plus adaptive particle correction rather than deterministic test-time tuning.

These applications show that “adaptive diffusion sampling” has expanded from inverse problems with explicit forward operators to settings where adaptation is driven by uncertainty calibration, semantic hierarchy, or particle-based posterior correction.

6. Theory, limitations, and recurring misconceptions

One recurring misconception is that ZADS is a single, settled algorithm. The literature instead supports a narrower claim: ZADS is a recurring zero-shot adaptation motif applied to several different reverse-time interventions. The official acronym appears in the MRI tuning work, whereas other papers explicitly describe the term as conceptual or note that it is not the paper’s official name (Alçalar et al., 11 Sep 2025, Alçalar et al., 2024, Wang et al., 10 Feb 2026).

A second misconception is that “zero-shot” means unguided. In every major formulation, adaptation is driven by some non-learned information available at inference time: measurements $\mathbf{W}\mathbf{D}_t\mathbf{W}^\top$9, operators $\|\mathbf{A}\mathbf{x}_0-\mathbf{y}\|_2^2$0 or $\|\mathbf{A}\mathbf{x}_0-\mathbf{y}\|_2^2$1, held-out k-space subsets, uncertainty estimates, or semantic stability matrices. The common restriction is not absence of adaptation, but absence of retraining of the generative prior.

A third misconception is that training-free implies cheap inference. AdaSense requires repeated posterior sampling at each adaptive step, with total cost approximately $\|\mathbf{A}\mathbf{x}_0-\mathbf{y}\|_2^2$2 forward passes per step (Elata et al., 2024). ZAPS performs several fine-tuning epochs over an unrolled sampler (Alçalar et al., 2024). The MRI ZADS method uses 25 steps, 15 CG iterations per step, and 10 fine-tuning epochs, yielding 250 NFEs in the reported protocol (Alçalar et al., 11 Sep 2025). The segmentation method has time complexity $\|\mathbf{A}\mathbf{x}_0-\mathbf{y}\|_2^2$3 and space complexity $\|\mathbf{A}\mathbf{x}_0-\mathbf{y}\|_2^2$4 (Nakhli et al., 14 Jun 2026). The zero-shot designation therefore refers to supervision and retraining, not to computational lightness.

Theoretical work on score mismatch clarifies why adaptation matters. For zero-shot conditional samplers, the asymptotic bias is proportional to accumulated mismatch between the target and training distributions, and the paper derives a bias-optimal zero-shot sampler for linear conditional models, BO-DDNM, that minimizes per-step mismatch. When $\|\mathbf{A}\mathbf{x}_0-\mathbf{y}\|_2^2$5, this sampler coincides with DDNM; for $\|\mathbf{A}\mathbf{x}_0-\mathbf{y}\|_2^2$6, it yields smaller asymptotic bias than DDNM under the stated analysis (Liang et al., 2024). This provides a formal basis for many practical ZADS mechanisms: they can be read as attempts to reduce mismatch by adjusting guidance, curvature, timestep placement, or data-consistency strength.

The limitations are likewise consistent across papers. Linearity is a central assumption in AdaSense and in several inverse-problem samplers, and extensions to nonlinear acquisition are explicitly described as future work (Elata et al., 2024). Gaussian likelihoods are the canonical model; hard data consistency is most appropriate in low-noise or noiseless regimes, whereas explicit likelihood-guided methods are preferred when measurement noise is non-negligible (Elata et al., 2024, Wang et al., 16 Jun 2025). Performance depends on the fidelity and domain match of the pre-trained prior, so mismatch in anatomy, coils, degradations, or image statistics can degrade both reconstruction and acquisition decisions (Elata et al., 2024, Wang et al., 10 Feb 2026). Adaptive timestep mechanisms also introduce their own sensitivities, such as the uncertainty threshold $\|\mathbf{A}\mathbf{x}_0-\mathbf{y}\|_2^2$7 in AdaDS and the PELT penalty $\|\mathbf{A}\mathbf{x}_0-\mathbf{y}\|_2^2$8 in segmentation (Wang et al., 10 Feb 2026, Nakhli et al., 14 Jun 2026).

Within those constraints, ZADS has become a useful descriptor for a broad class of methods that keep the diffusion prior fixed while adapting the inference path. Its major variants differ in what is adapted, but they share the same underlying ambition: to turn a pretrained diffusion model into a posterior-aware, schedule-aware, or uncertainty-aware sampler at test time.