Mixture-of-Gaussians-Guided Schedule Design for Brownian Bridge Diffusion Models

Michael Elad; Ron Levi

REVIEW 2 major objections 5 minor 27 references

Mixture-of-Gaussians-Guided Schedule Design for Brownian Bridge Diffusion Models

T0 review · 2 major / 5 minor · reviewed 2026-07-12 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Brownian-bridge diffusion schedules can be designed from first principles under a mixture-of-Gaussians prior, yielding universal MSE and Wasserstein optima that trade fidelity against perception.

desk verdict Solid MoG analysis of BBDM schedules with real proofs and a clean distortion–perception tradeoff; the “universal” real-image schedules are honest heuristics, not proven optima. read the letter →

arxiv 2607.03517 v1 pith:PKVBNXJ5 submitted 2026-07-03 cs.LG cs.CV

Ron Levi , Michael Elad This is my paper · ORCID

classification cs.LGcs.CV

keywords BrownianBridgeDiffusionModelsscheduledesignMixture-of-Gaussiansperception-distortiontradeoffinverseproblemsselected-labelapproximationWassersteinMSE

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Brownian Bridge Diffusion Models restore images by bridging the clean signal directly to the degraded observation rather than to pure noise, yet their bridge schedule has been chosen by heuristics. This paper supplies an analytical design method: under a mixture-of-Gaussians prior the ideal posterior and MMSE denoiser are closed-form, but the exact reverse chain loses global affinity. A selected-label surrogate freezes the component drawn from the measurement posterior, restoring an explicit reconstruction law that is mean-exact yet covariance-deficient. Matching that law to the true posterior produces two objectives—one Wasserstein, one MSE—that formalize the perception–distortion tradeoff and admit universal optima inside a bounded four-parameter family, independent of the degradation operator and prior. Controlled MoG experiments confirm the theory; FFHQ inpainting, deblurring and super-resolution show that the resulting MSE- and W2-oriented schedules deliver the expected fidelity-versus-perception tradeoff.

What carries the argument

Selected-label approximation: draw one mixture component from the measurement posterior and freeze it for the entire reverse chain, converting the non-affine MoG sampler into an explicit, diagonalizable Gaussian reconstruction law whose schedule dependence is fully analytic.

What would settle it

If, on a controlled shared-covariance MoG problem where the exact posterior is known, the schedule that minimises the analytic selected-label MSE (or W2) objective fails to minimise the corresponding empirical error of the oracle MoG BBDM chain, or if the FFHQ MSE-oriented schedule systematically loses both PSNR and SSIM to the default schedule, the central design claim is false.

Watch

Extended reading notes

Core claim

Under a Mixture-of-Gaussians prior the selected-label BBDM surrogate is component-wise mean-exact and covariance-deficient relative to the measurement posterior; the resulting matched-component Wasserstein and product-coupling MSE objectives therefore admit universal optima inside a bounded four-parameter schedule family that do not depend on the degradation, measurement noise or prior.

Load-bearing premise

The claim rests on the premise that freezing a single mixture label drawn from the measurement posterior faithfully tracks how the true reverse chain (and trained denoisers) respond to changes in the bridge schedule.

Editorial extensions

If this is right

Practitioners can adopt the two edge schedules (1,2,2,0.2) for fidelity and (2,1,0.2,2) for perception without fitting a high-dimensional MoG.
The same analytic pipeline applies verbatim to conditional DDIM, enabling direct schedule comparison between bridge and non-bridge samplers.
Any inverse problem whose degradation and prior admit a component-wise eigenbasis can reuse the closed-form reconstruction law for schedule search.
The explicit covariance-deficit formula quantifies how much posterior variance a given schedule can recover, turning schedule choice into a controllable perception–distortion dial.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the optima are degradation-independent, a single pair of schedules may transfer across an entire family of linear restoration tasks once a network is trained.
The covariance-shaping view suggests that other reverse-process hyperparameters (step size, stochasticity) could be optimised by the same matched-component criteria.
If a practical high-dimensional MoG estimator becomes available, the full objectives could replace the edge heuristics and yield problem-specific rather than universal schedules.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The useful core is an analytical pipeline for BBDM under a MoG prior: exact posterior and MMSE denoiser, then a selected-label freeze that restores affine unrolling, mean-exactness, and a covariance deficit relative to the measurement posterior. From that they get matched-component W2 and product-coupling MSE schedule objectives, plus a parallel conditional-DDIM appendix. That is a genuine extension of Benita et al.’s single-Gaussian program to mixtures and arbitrary linear H, not just a reparameterization of the bridge.

What they do well is the formal work. Appendices B–G actually derive the posteriors, the loss-of-affinity lemma, mean-exactness (Prop. 5.4), covariance deficit (Prop. 5.5), and a high-d self-consistency theorem for the freeze under shared covariance and linear separation. Controlled MoG experiments line up with the theory; MNIST shows the surrogate ranks schedules the same way a trained BBDM does. On FFHQ the two edge schedules produce the expected PSNR/SSIM vs FID/LPIPS tradeoff across blur, SR, and inpainting, and they beat default BBDM and a matched cDDIM baseline on the axes they target. Citations are appropriate; no circular fitting of schedules to FFHQ metrics.

Soft spots, in proportion: the four-parameter family is a convenience box, not the full schedule space. The MSE edge (1,2,2,0.2) follows cleanly from minimizing an upper bound on sampler variance; the W2 edge is the opposite direction on that same bound, not a proven minimizer of JW2 itself. Outside the shared-covariance regime of Theorem G.2 the freeze is an approximation, so for real images they correctly fall back to MoG-free heuristics rather than claiming fitted high-d MoG optima. No code release is a practical annoyance, not a soundness hole. The stress-test is right that “universal optima independent of degradation and prior” is overstated for the W2 edge and for real images; the paper’s own experiments still show the heuristics work as directional tools.

This is for people who care about schedule design and analytical reverse dynamics in diffusion inverse problems. It deserves a serious referee. I would engage with it and cite the surrogate + objectives when discussing BBDM or conditional DDIM schedules.

Referee Report

2 major / 5 minor

Summary. The paper develops an analytical framework for schedule design in Brownian Bridge Diffusion Models (BBDM) for linear inverse problems under a Mixture-of-Gaussians (MoG) prior. It derives the exact posterior p(x0|xs,y) and MMSE denoiser (Lemmas 3.1–3.2, Eq. 19), shows that the oracle MoG reverse chain loses global affinity (Lemma 4.1), and introduces a selected-label surrogate that freezes a measurement-posterior component label. The surrogate is proven mean-exact (Prop. 5.4) and covariance-deficient (Prop. 5.5) relative to the true measurement posterior, yielding an explicit reconstruction law (Cor. 5.3). From this, matched-component Wasserstein (JW2_MOG, Cor. 6.4) and product-coupling MSE (JMSE_MOG, Cor. 6.5) objectives are obtained, exposing a distortion–perception tradeoff. Inside a bounded four-parameter schedule family, MoG-free edge rules are derived (App. F): MSE upper-bound minimizer (1,2,2,0.2) and opposite W2 heuristic (2,1,0.2,2), claimed independent of degradation, noise and prior. Controlled MoG experiments, MNIST surrogate-to-trained transfer, and FFHQ restoration tasks (blur, SR, inpainting) validate the analysis and the resulting schedules against default BBDM and conditional DDIM.

Significance. If the analysis holds, the work supplies the first closed-form, method-agnostic schedule-design criteria for BBDM that go beyond single-Gaussian Fourier-shared-basis settings (Benita et al.). The exact MoG posterior, selected-label unrolling, mean-exactness/covariance-deficit proofs, and self-consistency theorem (Thm. G.2) are rigorous contributions; the controlled MoG experiments fully align theory with practice, and the FFHQ results demonstrate a usable distortion–perception tradeoff via two fixed schedules. These strengths—machine-checkable derivations in Apps. B–G, falsifiable schedule rankings, and reproducible edge heuristics—make the paper a solid advance for analytical diffusion inference and inverse-problem sampling, even if the real-data schedules remain heuristic.

major comments (2)

Abstract and §1 claim that the work “proves the existence of universal schedules for both [MSE and W2] that are independent of the degradation and prior.” Corollaries 6.4–6.5 and App. F establish this rigorously only for the MSE upper bound on σ²_BBDM (minimized by driving ρ1 o0, yielding the edge (α,β,c,γ)=(1,2,2,0.2) independent of the λr,k). For JW2_MOG the same appendix supplies only the opposite-edge heuristic that increases the variance upper bound; it is not shown to minimize the actual matched-component W2 objective. The abstract therefore overstates the proven scope. Soften the claim to “universal MSE schedule and a practical W2 heuristic” or supply a proof that the edge is optimal for JW2 itself.
The load-bearing selected-label approximation (Sec. 5, Cor. 5.3) is rigorously self-consistent only under the shared-covariance, linearly-separated high-dimensional regime of Theorem G.2. Outside that regime the surrogate objectives may mis-rank schedules relative to the oracle MoG chain or a trained denoiser. While MNIST Spearman correlations are high and FFHQ shows the expected tradeoff, the paper never quantifies surrogate-to-oracle discrepancy for non-shared or image-scale MoGs; the MoG-free edges used on FFHQ therefore rest on an unvalidated transfer. A short ablation (or explicit caveat) measuring schedule ranking fidelity under non-shared covariances would strengthen the central practical claim.

minor comments (5)

Fig. 1 caption and §6: the three schedules are plotted for S=200, yet many analytic comparisons use S=20; a brief note that the edge shapes are S-independent would avoid confusion.
Notation: the same symbol δs is used both for the bridge variance schedule and (in places) for the reverse-step variance; a subscript or distinct symbol would improve readability.
Table 1 (MNIST): NLL is reported from an external classifier; state the exact checkpoint and whether the metric is averaged over the full test set or a subset.
App. H (conditional DDIM): the endpoint idealization ᾱS=0 is used for mean-exactness; a short remark on the practical error when ᾱS is only “close to zero” would be useful.
Typos: “MOG” vs “MoG” inconsistency; “ass” for “as” in a few places (e.g., near Eq. 1); “super-j/vit-base-mnist” should be cited with a stable identifier.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: schedule objectives and MoG-free edge rules are derived from the selected-label surrogate and a variance bound, then evaluated a priori on data.

full rationale

The derivation chain is self-contained and does not reduce claimed results to their inputs by construction. The exact MoG posterior and MMSE denoiser (Section 3, Appendix B) are standard Gaussian conditioning; the selected-label surrogate (Section 5) is an explicit approximation whose mean-exactness and covariance deficit (Props. 5.4–5.5) are proved from the fixed-label recursion, not assumed. The JW2 and JMSE objectives (Corollaries 6.4–6.5) compare that surrogate law to the true measurement posterior under stated couplings; they are not fitted to reconstruction metrics. The “universal” MSE-oriented edge (α,β,c,γ)=(1,2,2,0.2) follows from monotonicity of an upper bound on σ²_BBDM in ρ1 (Appendix F.3), which is independent of the λr,k by algebra of the bound—not by fitting free parameters to data. The W2 edge is openly labeled a heuristic (F.4: “not a certificate of a universal W2 optimum”). FFHQ/MNIST schedules are fixed from these a priori rules (or surrogate optimization on MNIST) and then measured; there is no fit-then-predict loop on the same quantities. Citations to Benita et al. (2025; 2026) situate the single-Gaussian program being extended; they are not load-bearing uniqueness or ansatz imports for the MoG proofs. Overclaim that the abstract “proves” universal optima for both objectives (while W2 is only a bound-direction heuristic) is a correctness/strength issue, not circularity. No self-definitional step, fitted-input-as-prediction, or renaming of a known empirical pattern as a first-principles derivation was found.

Assumptions & free parameters 3 free parameters · 6 assumptions · 2 invented entities

The central claims rest on the BBDM bridge construction (prior literature), a linear Gaussian measurement model, an MoG prior that makes posteriors closed-form, the selected-label freeze that restores affinity, and a bounded four-parameter schedule family whose edge optima are read off a variance upper bound. No new physical entities are postulated; the free parameters are the schedule-box bounds and the usual diffusion hyperparameters.

free parameters (3)

schedule family bounds (α,β,c,γ) ∈ [1,2]×[1,2]×[0.2,2]×[0.2,2]
Hand-chosen search box that keeps bridges non-degenerate; edge optima (1,2,2,0.2) and (2,1,0.2,2) are selected inside this box, not over all valid schedules.
number of reverse steps S
Discrete sampling budget (e.g. S=20 toy, S=200 FFHQ) that enters the unrolled variance formulas and experimental comparisons.
default reference schedule δ_ref_s = 2(m_s − m_s²) and default (α,β,c,γ)=(1,1,0.5,1)
Baseline inherited from original BBDM heuristics; used as comparison point rather than derived optimum.

assumptions (6)

domain assumption Clean signals follow a finite Mixture-of-Gaussians prior with positive-definite component covariances.
Section 2.3 / Eq. (8); enables closed-form component posteriors and MMSE denoiser.
domain assumption Observations obey a linear Gaussian model y|x0 ~ N(Hx0, σ_y² I) with known H.
Section 2.3 / Eq. (7); standard inverse-problem assumption used throughout the posterior derivations.
domain assumption BBDM forward marginal and reverse update coefficients (a_s,b_s,c_s,σ_s) as in Li et al. (2023).
Section 2.1–2.2; the analysis designs the schedule inside this fixed sampler family.
ad hoc to paper Selected-label freeze: J ~ γ·|y is drawn once and held fixed for all reverse steps.
Section 5; the device that restores global affinity and yields the unrolled Gaussian law.
standard math Standard Gaussian conditioning identities and closed-form W2 between Gaussians (Dowson & Landau).
Appendix A and Lemma 6.3; classical facts used without re-proof.
ad hoc to paper Shared-covariance, linear measurement separation of order d for selected-label self-consistency.
Assumption G.1 and Theorem G.2; required for the high-probability responsibility concentration argument.

invented entities (2)

Selected-label BBDM surrogate reconstruction law p_Select(x̂0|y) independent evidence
purpose: Tractable mixture of Gaussians that approximates the oracle MoG reverse chain and exposes schedule dependence of mean and covariance.
Introduced in Section 5 / Corollary 5.3; mean-exact and covariance-deficient by Propositions 5.4–5.5.
Matched-component Wasserstein upper-bound objective J_W2_MOG and product-coupling MSE objective J_MSE_MOG independent evidence
purpose: Two complementary scalar criteria for optimizing the bridge schedule toward perception vs. distortion.
Corollaries 6.4–6.5; derived from the surrogate-versus-true-posterior comparison.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mixture-of-Gaussians-Guided Schedule Design for Brownian Bridge Diffusion Models." pith.science (2026). https://pith.science/paper/PKVBNXJ5

@misc{pith2026260703517,
  author       = {Pith},
  title        = {Pith review of: Mixture-of-Gaussians-Guided Schedule Design for Brownian Bridge Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PKVBNXJ5}},
  note         = {Machine review of arXiv:2607.03517}
}

read the original abstract

Brownian Bridge Diffusion Models (BBDM) offer an appealing framework for image restoration and inverse problems by constructing a stochastic bridge from the clean signal directly to the degraded observation, rather than to pure noise. Despite their promise, the choice of bridge schedule is typically inherited from heuristics, and a principled analytical framework for schedule design has been lacking. In this work, we develop such a framework by offering a novel analysis of BBDM reverse dynamics under a Mixture-of-Gaussians (MoG) prior. This setting yields a closed-form ideal posterior and a corresponding MMSE denoiser, while the BBDM-induced reconstruction law is captured analytically through a tractable surrogate. Building on these expressions, we formulate two complementary schedule-design objectives: a Wasserstein criterion targeting perceptual quality and an MSE criterion targeting reconstruction fidelity. Our work exposes an inherent tradeoff between the two and proves the existence of universal schedules for both that are independent of the degradation and prior. Extensive experiments on controlled MoG settings confirm full alignment between theory and practice, and experiments on the FFHQ dataset across inpainting, deblurring, and super-resolution tasks validate the practical value of our schedule-design criteria.

Figures

Figures reproduced from arXiv: 2607.03517 by the authors.

**Figure 2.** Shared-covariance toy experiment for schedules optimized with [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

**Figure 3.** Step-wise analytic comparison between conditional DDIM and BBDM schedules in the shared [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Dependency graph for the exact MoG posterior derivation. [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]

**Figure 5.** Figure 5: High-dimensional validation of the frozen-selection approximation in the controlled shared [PITH_FULL_IMAGE:figures/full_fig_p060_5.png]

**Figure 6.** Figure 6: One-dimensional MNIST schedule sweeps for the baseline blur setting [PITH_FULL_IMAGE:figures/full_fig_p061_6.png]

**Figure 7.** Figure 7: Effect of prior richness in the frozen selected-label surrogate for MNIST deblurring. The number [PITH_FULL_IMAGE:figures/full_fig_p062_7.png]

**Figure 8.** Figure 8: Qualitative FFHQ blur reconstructions for [PITH_FULL_IMAGE:figures/full_fig_p064_8.png]

**Figure 9.** Figure 9: Qualitative FFHQ 8× super-resolution reconstructions with σy = 0.1. Columns show the original image, degraded observation, MSE-oriented reconstruction, default reconstruction, and W2-oriented reconstruction. 65 [PITH_FULL_IMAGE:figures/full_fig_p065_9.png]

**Figure 10.** Figure 10: Qualitative FFHQ distributed inpainting reconstructions for [PITH_FULL_IMAGE:figures/full_fig_p066_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 3 linked inside Pith

[1]

Albergo and Nicholas M

Michael S. Albergo and Nicholas M. Boffi and Eric Vanden-Eijnden , title =. arXiv preprint arXiv:2303.08797 , year =

arXiv
[2]

arXiv preprint arXiv:2502.00180 , year =

Roi Benita and Michael Elad and Joseph Keshet , title =. arXiv preprint arXiv:2502.00180 , year =

arXiv
[3]

arXiv preprint arXiv:2602.07715 , year =

Roi Benita and Michael Elad and Joseph Keshet , title =. arXiv preprint arXiv:2602.07715 , year =

arXiv
[4]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

Yochai Blau and Tomer Michaeli , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =
[5]

Journal of Mathematical Imaging and Vision , volume =

Nicolas Bonneel and Julien Rabin and Gabriel Peyr\'e and Hanspeter Pfister , title =. Journal of Mathematical Imaging and Vision , volume =
[6]

McCann and Marc L

Hyungjin Chung and Jeongsol Kim and Michael T. McCann and Marc L. Klasky and Jong Chul Ye , title =. International Conference on Learning Representations (ICLR) , year =
[7]

Advances in Neural Information Processing Systems (NeurIPS) , year =

Valentin De Bortoli and James Thornton and Jeremy Heng and Arnaud Doucet , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =
[8]

D. C. Dowson and B. V. Landau , title =. Journal of Multivariate Analysis , volume =

Show all 27 references

[9]

Advances in Neural Information Processing Systems (NeurIPS) , year =

Martin Heusel and Hubert Ramsauer and Thomas Unterthiner and Bernhard Nessler and Sepp Hochreiter , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =
[10]

Advances in Neural Information Processing Systems (NeurIPS) , year =

Jonathan Ho and Ajay Jain and Pieter Abbeel , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =
[11]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

Tero Karras and Samuli Laine and Timo Aila , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =
[12]

Advances in Neural Information Processing Systems (NeurIPS) , year =

Tero Karras and Miika Aittala and Timo Aila and Samuli Laine , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =
[13]

Advances in Neural Information Processing Systems (NeurIPS) , year =

Bahjat Kawar and Michael Elad and Stefano Ermon and Jiaming Song , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =
[14]

Proceedings of the IEEE , volume =

Yann LeCun and L\'eon Bottou and Yoshua Bengio and Patrick Haffner , title =. Proceedings of the IEEE , volume =
[15]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

Bo Li and Kaitao Xue and Bin Liu and Yu-Kun Lai , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =
[16]

Theodorou and Weili Nie and Anima Anandkumar , title =

Guan-Horng Liu and Arash Vahdat and De-An Huang and Evangelos A. Theodorou and Weili Nie and Anima Anandkumar , title =. International Conference on Machine Learning (ICML) , year =
[17]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

Andreas Lugmayr and Martin Danelljan and Andres Romero and Fisher Yu and Radu Timofte and Luc Van Gool , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =
[18]

International Conference on Machine Learning (ICML) , year =

Alexander Quinn Nichol and Prafulla Dhariwal , title =. International Conference on Machine Learning (ICML) , year =
[19]

Foundations and Trends in Machine Learning , volume =

Gabriel Peyr\'e and Marco Cuturi , title =. Foundations and Trends in Machine Learning , volume =
[20]

Electronic Communications in Probability , volume =

Mark Rudelson and Roman Vershynin , title =. Electronic Communications in Probability , volume =
[21]

Fleet and Mohammad Norouzi , title =

Chitwan Saharia and Jonathan Ho and William Chan and Tim Salimans and David J. Fleet and Mohammad Norouzi , title =. arXiv preprint arXiv:2104.07636 , year =

arXiv
[22]

Lee and Jonathan Ho and Tim Salimans and David J

Chitwan Saharia and William Chan and Huiwen Chang and Chris A. Lee and Jonathan Ho and Tim Salimans and David J. Fleet and Mohammad Norouzi , title =. ACM Transactions on Graphics , year =
[23]

International Conference on Learning Representations (ICLR) , year =

Jiaming Song and Chenlin Meng and Stefano Ermon , title =. International Conference on Learning Representations (ICLR) , year =
[24]

International Conference on Learning Representations (ICLR) , year =

Jiaming Song and Arash Vahdat and Morteza Mardani and Jan Kautz , title =. International Conference on Learning Representations (ICLR) , year =
[25]

Bovik and Hamid R

Zhou Wang and Alan C. Bovik and Hamid R. Sheikh and Eero P. Simoncelli , title =. IEEE Transactions on Image Processing , volume =
[26]

International Conference on Learning Representations (ICLR) , year =

Yinhuai Wang and Jiwen Yu and Jian Zhang , title =. International Conference on Learning Representations (ICLR) , year =
[27]

Efros and Eli Shechtman and Oliver Wang , title =

Richard Zhang and Phillip Isola and Alexei A. Efros and Eli Shechtman and Oliver Wang , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year =

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.

[1] [1]

Albergo and Nicholas M

Michael S. Albergo and Nicholas M. Boffi and Eric Vanden-Eijnden , title =. arXiv preprint arXiv:2303.08797 , year =

arXiv

[2] [2]

arXiv preprint arXiv:2502.00180 , year =

Roi Benita and Michael Elad and Joseph Keshet , title =. arXiv preprint arXiv:2502.00180 , year =

arXiv

[3] [3]

arXiv preprint arXiv:2602.07715 , year =

Roi Benita and Michael Elad and Joseph Keshet , title =. arXiv preprint arXiv:2602.07715 , year =

arXiv

[4] [4]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

Yochai Blau and Tomer Michaeli , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages =

[5] [5]

Journal of Mathematical Imaging and Vision , volume =

Nicolas Bonneel and Julien Rabin and Gabriel Peyr\'e and Hanspeter Pfister , title =. Journal of Mathematical Imaging and Vision , volume =

[6] [6]

McCann and Marc L

Hyungjin Chung and Jeongsol Kim and Michael T. McCann and Marc L. Klasky and Jong Chul Ye , title =. International Conference on Learning Representations (ICLR) , year =

[7] [7]

Advances in Neural Information Processing Systems (NeurIPS) , year =

Valentin De Bortoli and James Thornton and Jeremy Heng and Arnaud Doucet , title =. Advances in Neural Information Processing Systems (NeurIPS) , year =

[8] [8]

D. C. Dowson and B. V. Landau , title =. Journal of Multivariate Analysis , volume =