arxiv: 2604.21066 · v2 · submitted 2026-04-22 · 💻 cs.CV · cs.LG· stat.ME

Recognition: unknown

Optimizing Diffusion Priors in Image Reconstruction from a Single Observation

Frederic Wang , Katherine L. Bouman

Authors on Pith no claims yet

Pith reviewed 2026-05-10 00:17 UTC · model grok-4.3

classification 💻 cs.CV cs.LGstat.ME

keywords diffusion modelsproduct of expertsBayesian evidence maximizationinverse imaging problemsprior adaptationsingle-shot tuningblack hole imagingimage deblurring

0 comments

The pith

A product-of-experts combination of diffusion priors can be tuned from a single observation by maximizing Bayesian evidence.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows how to adapt diffusion priors for inverse problems using only one observation. It combines multiple existing priors into a product-of-experts form and selects the weighting exponents that maximize the Bayesian evidence of the observation. This sidesteps the requirement for large numbers of varied measurements that current finetuning methods need. The resulting prior supports posterior sampling that is less biased by the original training data. Experiments on black hole imaging and text-conditioned deblurring confirm that evidence maximization often selects combinations extending beyond any single original prior.

Core claim

The authors claim that by forming a product-of-experts prior from several diffusion models and optimizing the exponents through maximization of the Bayesian evidence computed on a lone observation, one obtains a more generalizable prior suitable for reconstructing images in challenging inverse problems where the ground-truth distribution is unknown.

What carries the argument

The product-of-experts prior whose exponents are chosen to maximize the evidence of a single measurement under the combined model.

If this is right

Posterior sampling becomes possible from both tempered and combined diffusion models.
The approach produces priors that extend beyond those trained on any single dataset.
Image reconstructions gain trustworthiness in settings like black hole imaging with unknown true priors.
Text-conditioned deblurring benefits from the exponent-weighted combination.
No large collection of observations with different forward operators is required for prior tuning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The technique may extend to selecting among other types of generative priors in data-scarce Bayesian inference.
Evidence-based exponent tuning could reduce overfitting risks when adapting models to new domains with minimal data.
It opens the possibility of dynamically weighting priors based on how well they explain the available observation.

Load-bearing premise

That the exponent values maximizing the Bayesian evidence on one observation will produce a prior whose induced posterior reliably approximates the true solution distribution.

What would settle it

A test where synthetic data is generated from a known ground-truth distribution, a single observation is created, the exponents are optimized, and the sampled posteriors are compared to the known true distribution for accuracy and calibration.

Figures

Figures reproduced from arXiv: 2604.21066 by Frederic Wang, Katherine L. Bouman.

**Figure 1.** Figure 1: Top: Expectation-maximization iterations for the exponents a, b of the 1000- dimensional product prior ∝ p(x) a q(x) b . Here, p(x) represents a narrow Gaussian prior while q represents a wider Gaussian prior. Paths for an in-distribution, slightly outof-distribution, and fully out-of-distribution measurement (with respect to prior p) with different initializations are shown. The analytic evidence is also… view at source ↗

**Figure 2.** Figure 2: Unconditional samples of the product of the GRMHD and space priors. Ground Truth 1 Evidence of prior exponents from a GRMHD ground truth image Ground Truth EM Trajectories overlayed on evidence from a GRMHD ground truth image 2 posterior samples from max 1 max 2 [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Figure 3: Evidence fields for measurements of two in-distribution GRMHD images. Left: Because the ring size is unusually large relative to the training set, the evidence is maximized by tempering the GRMHD prior and combining it with the more general space prior. Posterior samples at the evidence-maximizing exponents are shown. Right: Evidence field and EM trajectories for a highly typical GRMHD image. The evidence … view at source ↗

**Figure 4.** Figure 4: Evidence fields for measurements for two slightly out-of-distribution images. Left: as the ground truth was drawn from p 0.5 GRMHDp 0.5 space, the evidence is maximized near this region, and posterior samples reconstruct the bright areas correctly up to translation invariance. Right: EM trajectories for an observation generated from a stretched GRMHD image. The evidence is maximized by relaxing the GRMHD p… view at source ↗

**Figure 5.** Figure 5: Evidence fields for real M87* black hole observations collected by the Event Horizon Telescope. Left: similar to previous atypical images, the evidence is maximized here by relaxing the GRMHD prior, either through tempering or by incorporating the more general space prior. Posterior samples show an increased range of possible structures compared to only using GRMHD as a prior (see supplement). Right: Evid… view at source ↗

**Figure 6.** Figure 6: Top: Exponents over many EM iterations for the golden corgi mixed-breed dog image. The EM converges to a large weight on the "golden retriever" and "corgi" priors and negative weight on the "poodle" prior. Bottom: Posterior and prior samples for iteration 1, which used the incorrect prior p(x | poodle), and iteration 10, which used the optimized prior. The reconstructions at iteration 10 are more semantica… view at source ↗

**Figure 7.** Figure 7: Left: Exponents over many EM iterations for a low exposure nighttime image of the Sacre Coeur, taken using the author’s smartphone and is unlikely to come from any training set. The EM converges to a large weight on the "cathedral" and "temple" priors, indicating that both priors are useful despite the ground truth not technically coming from either, as well as a large negative weight on the palace and cas… view at source ↗

**Figure 8.** Figure 8: Baseline finetuning method using a single measurement and a linear forward model. Likelihood p(y | x) is shown on the left. Starting from a Gaussian mixture prior at iteration 0, after many EM iterations the prior collapses onto the likelihood, indicating overfitting to the measurement. Diffusion EM Now we run the same ablation study but with the pre-trained GRMHD prior as our initialization, and the M87* … view at source ↗

**Figure 9.** Figure 9: Diffusion finetuning approaches from a single observation of the M87* black hole. As expected, the diffusion prior collapses onto the M87* likelihood and gets narrower at every iteration, to the point where all the unconditional samples at iteration 3 are nearly the same, making it an overfitted prior [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗

**Figure 10.** Figure 10: NRMSE for different p in the least squares weighting w (i) m,n ∝ 1 |g (i) m,n|p . The error metrics are good for p ≥ 2, with p = 2 being the best for the in-distribution and slightly OOD y, and p = 4 being the best for OOD y. GT for in-distribution y p=0 p=1 p=2 p=4 p=6 GT for slightly OOD y p=0 p=1 p=2 p=4 p=6 GT for OOD y p=0 p=1 p=2 p=4 p=6 GT for in-distribution y p=0 p=1 p=2 p=4 p=6 GT for slightly O… view at source ↗

**Figure 11.** Figure 11 [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗

**Figure 12.** Figure 12: EM trajectories of different initializations for the golden corgi experiment. All trajectories converge to a high weight on the "golden retriever" and "corgi" text conditions, and no weight on the "poodle" condition [PITH_FULL_IMAGE:figures/full_fig_p025_12.png] view at source ↗

**Figure 13.** Figure 13: , we use the Sacre Coeur example. For the same set of priors, all EM trajectories converge to the same exponents. Even for a slightly different set of priors, where the unconditional is missing ( [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗

read the original abstract

While diffusion priors generate high-quality posterior samples across many inverse problems, they are often trained on limited training sets or purely simulated data, thus inheriting the errors and biases of these underlying sources. Current approaches to finetuning diffusion models rely on a large number of observations with varying forward operators, which can be difficult to collect for many applications, and thus lead to overfitting when the measurement set is small. We propose a method for tuning a prior from only a single observation by combining existing diffusion priors into a single product-of-experts prior and identifying the exponents that maximize the Bayesian evidence. We validate our method on real-world inverse problems, including black hole imaging, where the true prior is unknown a priori, and image deblurring with text-conditioned priors. We find that the evidence is often maximized by priors that extend beyond those trained on a single dataset. By generalizing the prior through exponent weighting, our approach enables posterior sampling from both tempered and combined diffusion models, yielding more flexible priors that improve the trustworthiness of the resulting posterior image distribution.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a workable way to blend diffusion priors for single-observation inverse problems by exponent weighting and evidence maximization, but the single-data tuning step risks overfitting to noise or local features in high dimensions.

read the letter

The core idea is to take several pre-trained diffusion priors, combine them as a product of experts with learnable exponents, and pick those exponents by maximizing the Bayesian evidence on the one available observation. This sidesteps the usual need for many measurements to fine-tune a prior, which matters for settings like black-hole imaging where extra views are expensive or impossible to get. The abstract notes that the optimal exponents often pull in models trained on different datasets, producing more flexible priors than any single one alone, and they report tests on real deblurring and black-hole data that show improved posterior trustworthiness. That practical framing is the paper's clearest strength; it targets a constraint that shows up often in applied imaging work. The main soft spot is the evidence-maximization step itself. With only one high-dimensional observation, the marginal likelihood estimate can favor exponents that match noise realizations or dataset-specific artifacts rather than the underlying distribution. Monte-Carlo or variational approximations to that integral are already high-variance in image space, so stability across independent observations from the same forward model becomes important to check. The abstract does not spell out those controls or give quantitative results, so it is hard to judge how well the method holds up. Readers working on diffusion models for inverse problems with limited data will find the technique relevant and worth trying. The idea is new enough and the problem real enough that the paper should go to peer review for a closer look at the implementation and experiments.

Referee Report

3 major / 2 minor

Summary. The paper claims to introduce a method for tuning diffusion priors from a single observation by forming a product-of-experts combination of existing diffusion models and selecting the combination exponents that maximize the Bayesian evidence p(y|alpha). It asserts that this yields more flexible priors (including tempered and multi-dataset combinations) that improve the trustworthiness of posterior samples in inverse problems, with validation on black-hole imaging (where the true prior is unknown) and text-conditioned image deblurring.

Significance. If the central claim holds, the work would be significant for data-scarce inverse problems in computer vision and scientific imaging, where collecting multiple observations or large training sets is impractical. It offers a principled, evidence-based route to adapt and combine pre-trained diffusion priors without additional data collection, potentially reducing biases from limited training distributions while enabling posterior sampling from hybrid models.

major comments (3)

[Abstract / Method] The core procedure for estimating the marginal likelihood p(y | alpha) = integral p(y|x) p(x|alpha) dx in high-dimensional image space is load-bearing for the claim, yet the abstract supplies no equations, approximation scheme (e.g., variational or Monte-Carlo), or variance analysis; high-variance estimates from a single y can favor exponents that fit noise realizations rather than the data distribution.
[Experiments] Validation on black-hole imaging and deblurring does not include a stability test: whether the argmax_alpha remains consistent or optimal when a second independent observation y' drawn from the same forward model is substituted. This directly tests the generalizability of the evidence-maximized prior and is required to address the overfitting risk.
[Abstract / Results] The claim that 'evidence is often maximized by priors that extend beyond those trained on a single dataset' is central but unsupported by any reported quantitative metrics, baselines, or error bars in the abstract; without these, it is impossible to judge whether the improvement in posterior trustworthiness is statistically meaningful or merely anecdotal.

minor comments (2)

[Method] Clarify the precise definition of the product-of-experts prior (including how the exponents enter the joint density) and any normalization constants that may be required for proper Bayesian evidence computation.
[Abstract] The abstract mentions 'tempered and combined diffusion models' without a brief reference or equation; a short parenthetical or citation would improve readability for readers unfamiliar with tempering in diffusion literature.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments, which highlight important aspects of clarity, validation, and evidence presentation. We address each major comment below and will revise the manuscript accordingly to strengthen the work while preserving its core contributions.

read point-by-point responses

Referee: [Abstract / Method] The core procedure for estimating the marginal likelihood p(y | alpha) = integral p(y|x) p(x|alpha) dx in high-dimensional image space is load-bearing for the claim, yet the abstract supplies no equations, approximation scheme (e.g., variational or Monte-Carlo), or variance analysis; high-variance estimates from a single y can favor exponents that fit noise realizations rather than the data distribution.

Authors: We agree that the abstract would benefit from greater specificity on the marginal likelihood estimation. The method section describes a Monte Carlo approximation to p(y|alpha) that leverages samples drawn from the product-of-experts diffusion prior, together with importance weighting to reduce variance. We will revise the abstract to include the defining integral, a concise statement of the Monte Carlo scheme, and a brief note on variance control. This addition will make the procedure transparent without changing the technical approach. revision: yes
Referee: [Experiments] Validation on black-hole imaging and deblurring does not include a stability test: whether the argmax_alpha remains consistent or optimal when a second independent observation y' drawn from the same forward model is substituted. This directly tests the generalizability of the evidence-maximized prior and is required to address the overfitting risk.

Authors: We acknowledge the value of an explicit stability test for assessing overfitting risk. In the black-hole imaging experiment only a single real observation exists, precluding a second independent y'; this is the motivating data-scarce regime. For the synthetic deblurring experiments we can generate additional independent observations under the same forward model. We will add a stability analysis in the revised manuscript that reports the consistency of the selected exponents across multiple y' draws and quantifies any variation in downstream posterior metrics. This will directly address the generalizability concern for the cases where it is feasible. revision: partial
Referee: [Abstract / Results] The claim that 'evidence is often maximized by priors that extend beyond those trained on a single dataset' is central but unsupported by any reported quantitative metrics, baselines, or error bars in the abstract; without these, it is impossible to judge whether the improvement in posterior trustworthiness is statistically meaningful or merely anecdotal.

Authors: We agree that the abstract should convey the quantitative support for this claim. The results section already contains comparisons against single-dataset baselines together with error bars obtained from repeated posterior sampling runs. We will revise the abstract to include concise quantitative statements (e.g., average improvement in a trustworthiness metric with standard deviation) that summarize the evidence-maximization findings. This will allow readers to assess statistical meaningfulness directly from the abstract while retaining the full details in the main text. revision: yes

Circularity Check

1 steps flagged

Exponents for product-of-experts prior are fitted by maximizing evidence on the single observation y itself

specific steps

fitted input called prediction [Abstract]
"We propose a method for tuning a prior from only a single observation by combining existing diffusion priors into a single product-of-experts prior and identifying the exponents that maximize the Bayesian evidence."

The exponents are identified by maximizing evidence on the single observation y; the resulting prior is then applied to reconstruct from that same y. The optimization step therefore fits the prior parameters directly to the target data, making the 'tuned prior' and its posterior samples statistically dependent on y by construction rather than independently derived.

full rationale

The paper's central procedure selects the exponents alpha by maximizing Bayesian evidence p(y|alpha) on the identical single observation y that is later used for posterior sampling p(x|y,alpha). This matches the fitted-input-called-prediction pattern: the prior is tuned directly to the target data rather than derived from independent sources or held-out statistics. While the abstract presents this as a generalizable tuning method, the load-bearing step reduces the prior choice to a fit on y, creating partial circularity in the claim that the resulting posterior is trustworthy. No other self-citation or self-definitional reductions are evident from the provided text.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The claim rests on the assumption that product-of-experts combinations of diffusion models remain valid priors and that evidence maximization on one observation produces a non-overfit, generalizable result.

free parameters (1)

exponents for each prior
The exponents are identified by maximizing the Bayesian evidence on the single observation.

axioms (1)

domain assumption Product-of-experts combination of diffusion priors forms a valid and flexible joint prior
Invoked when the paper states that combining existing priors into a single product-of-experts prior enables tuning from one observation.

pith-pipeline@v0.9.0 · 5480 in / 1240 out tokens · 44692 ms · 2026-05-10T00:17:36.098870+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

42 extracted references · 16 canonical work pages · 4 internal anchors

[1]

arXiv preprint arXiv:2403.08728 , year=

Aali, A., Daras, G., Levac, B., Kumar, S., Dimakis, A.G., Tamir, J.I.: Ambient diffusionposteriorsampling:Solvinginverseproblemswithdiffusionmodelstrained on corrupted data. arXiv preprint arXiv:2403.08728 (2024)

work page arXiv 2024
[2]

Alam, M.T., Imam, R., Guizani, M., Karray, F.: Flare up your data: Diffusion- based augmentation method in astronomical imaging (2024)

2024
[3]

Advances in Neural Information Processing Systems37, 19447–19471 (2024)

Bai, W., Wang, Y., Chen, W., Sun, H.: An expectation-maximization algorithm for training clean diffusion models from corrupted observations. Advances in Neural Information Processing Systems37, 19447–19471 (2024)

2024
[4]

The Astrophysical Journal980(1), 108 (2025)

Barco, G.M., Adam, A., Stone, C., Hezaveh, Y., Perreault-Levasseur, L.: Tack- ling the problem of distributional shifts: Correcting misspecified, high-dimensional data-driven priors for inverse problems. The Astrophysical Journal980(1), 108 (2025)

2025
[5]

The Astrophysical Journal894(1), 31 (2020)

Blackburn, L., Pesce, D.W., Johnson, M.D., Wielgus, M., Chael, A.A., Christian, P., Doeleman, S.S.: Closure statistics in interferometric data. The Astrophysical Journal894(1), 31 (2020)

2020
[6]

arXiv preprint arXiv:2310.06721 , year=

Boys, B., Girolami, M., Pidstrigach, J., Reich, S., Mosca, A., Akyildiz, O.D.: Tweedie moment projected diffusions for inverse problems. arXiv preprint arXiv:2310.06721 (2023)

work page arXiv 2023
[7]

Diffusion Posterior Sampling for General Noisy Inverse Problems

Chung, H., Kim, J., Mccann, M.T., Klasky, M.L., Ye, J.C.: Diffusion posterior sam- pling for general noisy inverse problems. arXiv preprint arXiv:2209.14687 (2022)

work page internal anchor Pith review arXiv 2022
[8]

arXiv preprint arXiv:2310.01110 (2023)

Chung, H., Ye, J.C., Milanfar, P., Delbracio, M.: Prompt-tuning latent diffusion models for inverse problems. arXiv preprint arXiv:2310.01110 (2023)

work page arXiv 2023
[9]

IEEE Transactions on Image Processing33, 3496–3507 (2024)

Coeurdoux, F., Dobigeon, N., Chainais, P.: Plug-and-play split gibbs sampler: em- bedding deep generative priors in bayesian inference. IEEE Transactions on Image Processing33, 3496–3507 (2024)

2024
[10]

arXiv preprint arXiv:2404.10177 (2024)

Daras, G., Dimakis, A.G., Daskalakis, C.: Consistent diffusion meets tweedie: Training exact ambient diffusion models with noisy data. arXiv preprint arXiv:2404.10177 (2024)

work page arXiv 2024
[11]

Advances in Neural Information Processing Systems36, 288–313 (2023)

Daras, G., Shah, K., Dagan, Y., Gollakota, A., Dimakis, A., Klivans, A.: Ambient diffusion: Learning clean distributions from corrupted data. Advances in Neural Information Processing Systems36, 288–313 (2023)

2023
[12]

Journal of the Royal Statistical Society Series B: Statistical Methodology68(3), 411–436 (2006)

Del Moral, P., Doucet, A., Jasra, A.: Sequential monte carlo samplers. Journal of the Royal Statistical Society Series B: Statistical Methodology68(3), 411–436 (2006)

2006
[13]

IEEE signal processing magazine29(6), 141–142 (2012)

Deng, L.: The mnist database of handwritten digit images for machine learning research [best of the web]. IEEE signal processing magazine29(6), 141–142 (2012)

2012
[14]

In: International confer- ence on machine learning

Du, Y., Durkan, C., Strudel, R., Tenenbaum, J.B., Dieleman, S., Fergus, R., Sohl- Dickstein, J., Doucet, A., Grathwohl, W.S.: Reduce, reuse, recycle: Compositional generation with energy-based diffusion models and mcmc. In: International confer- ence on machine learning. pp. 8489–8510. PMLR (2023)

2023
[15]

Wang and K.L

EHTC: First m87 eht results: Calibrated data (2019) 16 F. Wang and K.L. Bouman

2019
[16]

EHTC: First m87 event horizon telescope results. i. the shadow of the supermassive black hole. The Astrophysical Journal Letters875(1) (2019)

2019
[17]

The Astrophysical Journal975(2), 201 (2024)

Feng, B.T., Bouman, K.L., Freeman, W.T.: Event-horizon-scale imaging of m87* under different assumptions via deep generative image priors. The Astrophysical Journal975(2), 201 (2024)

2024
[18]

In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

Gao, A.F., Leong, O., Sun, H., Bouman, K.L.: Image reconstruction without ex- plicit priors. In: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5. IEEE (2023)

2023
[19]

Proteina: Scaling flow-based protein structure generative models.arXiv preprint arXiv:2503.00710, 2025

Geffner, T., Didi, K., Zhang, Z., Reidenbach, D., Cao, Z., Yim, J., Geiger, M., Dallago, C., Kucukbenli, E., Vahdat, A., et al.: Proteina: Scaling flow-based protein structure generative models. arXiv preprint arXiv:2503.00710 (2025)

work page arXiv 2025
[20]

Advances in neural information processing systems33, 6840–6851 (2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

2020
[21]

Classifier-Free Diffusion Guidance

Ho, J., Salimans, T.: Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598 (2022)

work page internal anchor Pith review arXiv 2022
[22]

Communications in Statistics-Simulation and Com- putation18(3), 1059–1076 (1989)

Hutchinson, M.F.: A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines. Communications in Statistics-Simulation and Com- putation18(3), 1059–1076 (1989)

1989
[23]

M., Liang, S., Alkhouri, I

Li, X., Kwon, S.M., Liang, S., Alkhouri, I.R., Ravishankar, S., Qu, Q.: Decoupled data consistency with diffusion purification for image restoration. arXiv preprint arXiv:2403.06054 (2024)

work page arXiv 2024
[24]

Universe8(2), 85 (2022)

Mizuno, Y.: Grmhd simulations and modeling for jet formation and acceleration region in agns. Universe8(2), 85 (2022)

2022
[25]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 10684–10695 (2022)

2022
[26]

Advances in Neural Information Processing Systems37, 87647–87682 (2024)

Rozet, F., Andry, G., Lanusse, F., Louppe, G.: Learning diffusion priors from obser- vations by expectation maximization. Advances in Neural Information Processing Systems37, 87647–87682 (2024)

2024
[27]

arXiv preprint arXiv:2503.02819 (2025)

Skreta, M., Akhound-Sadegh, T., Ohanesian, V., Bondesan, R., Aspuru-Guzik, A., Doucet, A., Brekelmans, R., Tong, A., Neklyudov, K.: Feynman-kac correc- tors in diffusion: Annealing, guidance, and product of experts. arXiv preprint arXiv:2503.02819 (2025)

work page arXiv 2025
[28]

arXiv preprint arXiv:2412.17762 (2024)

Skreta, M., Atanackovic, L., Bose, A.J., Tong, A., Neklyudov, K.: The super- position of diffusion models using the it\ˆ o density estimator. arXiv preprint arXiv:2412.17762 (2024)

work page arXiv 2024
[29]

In: International Conference on Learning Representations (2023)

Song, J., Vahdat, A., Mardani, M., Kautz, J.: Pseudoinverse-guided diffusion mod- els for inverse problems. In: International Conference on Learning Representations (2023)

2023
[30]

Advances in neural information processing systems34, 1415–1428 (2021)

Song, Y., Durkan, C., Murray, I., Ermon, S.: Maximum likelihood training of score- based diffusion models. Advances in neural information processing systems34, 1415–1428 (2021)

2021
[31]

Score-Based Generative Modeling through Stochastic Differential Equations

Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456 (2020)

work page internal anchor Pith review Pith/arXiv arXiv 2011
[32]

Acta numerica19, 451–559 (2010)

Stuart, A.M.: Inverse problems: a bayesian perspective. Acta numerica19, 451–559 (2010)

2010
[33]

Springer Nature (2017)

Thompson, A.R., Moran, J.M., Swenson, G.W.: Interferometry and synthesis in radio astronomy. Springer Nature (2017)

2017
[34]

Sample-efficient evidence estimation of score based priors for model selection

Wang, F., Bouman, K.L.: Sample-efficient evidence estimation of score based priors for model selection. arXiv preprint arXiv:2602.20549 (2026) Optimizing Diffusion Priors with a Single Observation 17

work page internal anchor Pith review Pith/arXiv arXiv 2026
[35]

Advances in Neural Infor- mation Processing Systems36, 31372–31403 (2023)

Wu, L., Trippe, B., Naesseth, C., Blei, D., Cunningham, J.P.: Practical and asymp- totically exact conditional sampling in diffusion models. Advances in Neural Infor- mation Processing Systems36, 31372–31403 (2023)

2023
[36]

Advances in Neural Information Processing Systems37, 118389–118427 (2024)

Wu, Z., Sun, Y., Chen, Y., Zhang, B., Yue, Y., Bouman, K.: Principled proba- bilistic imaging using diffusion models as plug-and-play priors. Advances in Neural Information Processing Systems37, 118389–118427 (2024)

2024
[37]

arXiv preprint arXiv:2510.01184 (2025) 14

Xu, Y., Wu, Y., Park, S., Zhou, Z., Tulsiani, S.: Temporal score rescaling for tem- perature sampling in diffusion and flow models. arXiv preprint arXiv:2510.01184 (2025)

work page arXiv 2025
[38]

In: Proceedings of the Computer Vision and Pattern Recognition Conference

Zhang, B., Chu, W., Berner, J., Meng, C., Anandkumar, A., Song, Y.: Improving diffusion inverse problem solving with decoupled noise annealing. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 20895–20905 (2025)

2025
[39]

arXiv preprint arXiv:2506.08894 , year=

Zhang, Y., Murtuza-Lanier, C., Li, Z., Du, Y., Wu, J.: Product of experts for visual generation. arXiv preprint arXiv:2506.08894 (2025)

work page arXiv 2025
[40]

arXiv preprint arXiv:2503.11043 , year=

Zheng, H., Chu, W., Zhang, B., Wu, Z., Wang, A., Feng, B.T., Zou, C., Sun, Y., Ko- vachki, N., Ross, Z.E., et al.: Inversebench: Benchmarking plug-and-play diffusion priors for inverse problems in physical sciences. arXiv preprint arXiv:2503.11043 (2025)

work page arXiv 2025
[41]

URL https://arxiv

Zhu, Y., Dou, Z., Zheng, H., Zhang, Y., Wu, Y.N., Gao, R.: Think twice before you act: Improving inverse problem solving with mcmc, 2024. URL https://arxiv. org/abs/2409.085511(2)

work page arXiv 2024
[42]

Optimizing Diffusion Priors in Image Reconstruction from a Single Observation

Zhu, Y., Zhang, K., Liang, J., Cao, J., Wen, B., Timofte, R., Van Gool, L.: De- noising diffusion models for plug-and-play image restoration. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 1219–1229 (2023) 18 F. Wang and K.L. Bouman Supplemental for "Optimizing Diffusion Priors in Image Reconstruction from a Sin...

2023