acceptodds
Under review as a conference paper at ICLR 2027

Mixed Poisson diffusion models and application to posterior sampling

Abstract

Diffusion models are a powerful framework for learning complex probability distributions. While widely used diffusion models for continuous data rely on Gaussian noising processes, we instead develop diffusion models based on Poisson noise. More precisely, we introduce and study a new mathematical framework, called Mixed Poisson Diffusion Models (MPDM), with noising dynamics driven by multivariate counting processes. Our methodology is motivated by Poisson inverse problems. The success of Gaussian diffusion models has led to their use as prior models in a growing number of posterior sampling methods for inverse problems. However, most of these approaches focus on Gaussian inverse problems, in which observations are generated through a deterministic forward operator corrupted by additive Gaussian noise. By contrast, Poisson inverse problems, where observations are discrete and Poisson distributed, remain comparatively underexplored in this literature. Building on our Poisson generative dynamics, we develop a natural framework for posterior sampling in Poisson inverse problems and derive several posterior sampling algorithms based on the proposed model. We evaluate our approach on realistic Poisson inverse problems and show that the resulting framework is well suited to this class of problems.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
diffusion models, generative models, inverse problems, posterior sampling, counting processes

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_iKu4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes to "poissonize" generative diffusion models for inverse problems: instead of considering diffusion models where the noising process is Gaussian to learn a prior and sampling the posterior, the authors propose to develop diffusion models based on Poisson. The main motivation is compatibility with the observation model: since Gaussian diffusion for posterior sampling may exploit the Gaussian structure of the measurement law, the authors aim to investigate whether the switching to Poisson can be beneficial for solving Poisson inverse problems. While the first part of the paper defines Poisson diffusion models and compares them with Gaussian ones, the second part uses this theory for Poisson inverse problems: the authors start with a simple inverse problem (Poisson denoising with forward operator aIda I_d), and then consider more challenging inverse problem to see whether a "Poissonian" approach is helpful.

Critical strengths

The paper is well written, and the mathematical content seems sound. Besides the general formulation of Poisson generative models, I believe it tackles an interesting problem: whether leveraging the inductive bias arising from compatibility with an observation model actually brings additional benefits in posterior sampling via diffusion-based generative models. The paper also includes experiments that help evaluate the Poissonian approach against the Gaussian one.

Critical weaknesses

The paper has the merit of raising several questions, but I personally did not find the answers fully satisfying. First, it shows that, for a simple Poisson-denoising example, Poisson denoising posterior sampling (PDPS) clearly outperforms the considered Gaussian-based samplers. This is interesting and seems to follow from Theorem 3, which holds, however, in very limited case, when the forward map A=aI\mathcal A = a I; see questions. For the other inverse problems considered in the paper, however, Theorem 3 is not available and additional approximations are needed. In such cases, there seems to be no clear advantage in using Poissonian generative models. These results leave unclear whether matching the generative dynamics with the measurement law provides an advantage beyond the simple setting of Theorem 3.

Questions for authors
  • In the denoising problem covered by Theorem 3, what is the relation between PDPS and the approach of Krull et al. (2024) in Image Denoising and the Generative Accumulation of Photons?
  • Why, in your opinion, does there not seem to be any clear advantage to using Poissianian rather Gaussian generative models for posterior sampling when the conditions of Theorem 3 are not satisfied? Do you think there is a way to study the mismatch in performance between Table 1 (where the Poissonian method is strikingly effective) and Table 2 (where there seems to be no clear indication that one approach is preferable to the other)?
  • Can Theorem 3 be extended beyond the case A=aI\mathcal A = a I?
4 comments

Thank you for your positive comments on the paper's presentation, and your interest in the question we tackle regarding the matching structure of the generative model and the likelihood of the Bayesian inverse problem. Thank you also for saying that Theorem 3 and the associated numerical experiments are interesting.

Our experiments do not establish that posterior sampling methods based on Poissonian generative models are systematically better than Gaussian ones for a general forward operator A\mathcal{A}. We provide a detailed answer to your question regarding extensions of Theorem 3 below, where we show that the "structural correspondence" of Theorem 3 is still useful for more general operators (specifically, circulant matrices). We think this constitutes a concrete example of the benefit of matching the generative dynamics with the Poisson inverse problems, beyond the simple A=aId\mathcal{A}=aI_d case.

First, regarding the general operator case, our objective is primarily to show that, beyond the favourable but admittedly limited scope of Theorem 3, several general-purpose Gaussian diffusion-based posterior sampling methods could be translated to the MPDM setting, and that the resulting algorithms were competitive for realistic Poisson inverse problems, which was a priori not obvious. The image reconstruction metrics reported in Table 2 do not necessarily reflect the quality of the approximate posterior distribution, but provide a simple way to verify that the proposed algorithms are viable. Ultimately, evaluation of posterior sample quality depends on the problem at hand. For instance, we report a calibration evaluation for the BrainWeb dataset in Appendix E.4 that assesses whether the sampling methods correctly detect lesions; we observe notably that despite its good image reconstruction metrics, the baseline (Gaussian) diffusion DPS is less effective compared to our MPDM-based methods (it tends to underestimate the presence of lesions). This is of course not proof that MPDM-based methods are systematically "better", but this shows that the image metrics of Table 2 are not the full story. Our point is that if someone trains a generative model with Poisson inverse problems in mind, opting for a Poissonian generative model rather than a classical Gaussian one is possible and potentially more principled.

Can Theorem 3 be extended beyond the case A=aId\mathcal{A}=aI_d ?

Thank you for this question. When A=(aij)\mathcal{A}=(a_{ij}) is a circulant matrix with non-negative coefficients, we can derive a specialized sampling algorithm that explicitly uses the correspondence between the Poissonian generative model and the likelihood. This case includes convolution-based blurring operators with periodic boundary conditions. We identified this extension after submitting the original manuscript. It shows that the case A=aId\mathcal{A}=aI_d is not the only case where the connection with Poisson generative models is mathematically explicit. We will include this extension in the revised paper, after conducting illustrative experiments.

The idea is to use latent-count augmentation, as in, eg, [7]. Define cij∼P(aijxj)c_{ij}\sim\mathcal{P}(a_{ij}x_j) such that yi=∑jcijy_i=\sum_j c_{ij}. Let Cj=∑icijC_j=\sum_i c_{ij}; since all column sums of A\mathcal{A} are equal to a constant aa, the augmented posterior (distribution of x,cx,c given yy) is

\pi(x)\times\prod_i 1\left(\sum_j c_{ij}=y_i\right) \times\prod_j e^{-ax_j}x_j^{C_j}\times \prod_{i,j}\frac{a_{ij}^{c_{ij}}}{c_{ij}!},$$ Given $x$ and $y_i$, $c_{i,...}$ follows a multinomial distribution $$c_{i,\ldots} \mid x,y_i \sim \text{Multinomial}\left( y_i, \left[ \frac{a_{ij}x_j}{(Ax)_i} \right]^{j=1,...,d} \right)$$ which can be sampled exactly. This part is already in [7]. The other conditional, the distribution of $x$ given $y,c$, is $$\pi(d x\vert c,y) \propto \pi(dx)\prod_j e^{-ax_j}x_j^{C_j}$$ This is the posterior distribution $\pi^C$ of $X\sim\pi$ given a Poisson observation $C\sim\mathcal{P}(aX)$! So Theorem 3 says \$ K_t^{\pi^C}(z)=K_{t+a}^{\pi}(z+C)\$ and we can sample from this conditional by simulating the associated MPP. This yields an approximate Gibbs sampler for the original posterior: alternate between sampling the latent counts from their multinomial conditional and an MPP simulation step. Compared with the general-purpose algorithms described in our paper, this approach avoids the denoiser/score approximation used in DPS and the Langevin MCMC likelihood step used in PnP-DM and DAPS. The latent-count augmentation itself is already in [7], where the multinomial step is the same. However, they use additional data augmentation and MCMC for the second step, while MPDM tackles it directly. [7] Elhadji Cisse Faye, Mame Diarra Fall, Nicolas Dobigeon, and Eric Barat. Bregman geometry-aware split gibbs sampling for bayesian poisson inverse problems, 2026.

A separate extension of Theorem 3 is available when A=diag(a1,...,ad)\mathcal{A}=\mathrm{diag}(a_1,...,a_d) is a diagonal matrix; however it requires to extend the definition of the denoiser to a vector-valued time t=(t1,...,td)\mathbf{t}=(t_1,...,t_d). Then, for 1≤i≤d1\leq i \leq d, Ktπy(z)=Kt+aπ(z+y)K_{\mathbf{t}}^{\pi^y}(z)= K_{\mathbf{t}+\mathbf{a}}^{\pi}(z+y) ; with a=(a1,...,ad)\mathbf{a}=(a_1,...,a_d). So, theoretically, if one were to train an anisotropic denoiser, where each pixel is noised at a different level tit_i zi∼P(tixi)z_i\sim\mathcal{P}(t_ix_i), instead of the isotropic one where ti=tt_i=t for all ii, the exact conditional denoiser would be available without additional approximation. A similar idea is explored in [8] for Gaussian diffusion/Gaussian inverse problems. We will include a remark after the proof of Theorem 3; but we think this generalization lies beyond the scope of our manuscript, as it requires non-standard denoiser training.

In the denoising problem covered by Theorem 3, what is the relation between PDPS and the approach of Krull et al. (2024) in Image Denoising and the Generative Accumulation of Photons?

Thank you for pointing out this paper; it is an interesting related work, and we will mention it in the revised manuscript. While their approach and construction differ from both ours and the previous Poissonian generative models we identified in the paper [1,2], they also define and train a Poisson denoiser. They describe their sampling procedure as "predicting the next photon from an intermediary state"; this is reminiscent of the jump mechanism of counting processes. Their goal is to obtain posterior samples from a corrupted Poisson observation, which is also the objective of Theorem 3. However, they do not write a relationship between prior and posterior denoisers; they directly apply their photon arrival scheme to the observation. By contrast, PDPS (our algorithm that follows from Theorem 3) samples a mixed Poisson process from t=0,Zt=0t=0, Z_t=0 (the initial black image) and incorporates the observation into the denoiser.

[1] Sagnik Bhattacharya, Abhiram Gorle, Ahsan Bilal, Connor Ding, Amit Kumar Singh Yadav, and Tsachy Weissman. Itdpdm: Information-theoretic discrete poisson diffusion model. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, editors, Advances in Neural Information Processing Systems, volume 38, Main Conference, pages 10860–10915. Curran Associates, Inc., 2025.

[2] Tianqi Chen and Mingyuan Zhou. Learning to Jump: Thinning and Thickening Latent Counts for Generative Modeling. In Proceedings of the 40th International Conference on Machine Learning, pages 5367–5382. PMLR, July 2023. ISSN: 2640-3498

[8] Abbas Mammadov, Ozgur Kara, Kaan Oktay, Iskander Azangulov, Adil Kaan Akan, Hyungjin Chung, James Matthew Rehg, and Yee Whye Teh. Exact posterior score estimation for solving linear inverse problems, 2026.

I thank the authors for their explanation. The proposed extension to circulant operators is a valuable addition.

What remains unclear to me is the broader significance of the analysis. Why should we "poissonize" generative diffusion models for Poisson inverse problems? I would find the motivation stronger if the paper showed a concrete limitation of Gaussian diffusion models that the proposed approach can overcome in practice. However, the reconstruction results are mixed. I still think it is interesting that the paper gives a rigorous formulation of Poisson generative diffusion models and explores their application to inverse problems. But I also think that the paper in its current form gives practitioners little guidance on when to choose a Poisson generative model over a Gaussian diffusion model paired with the same Poisson likelihood.

We thank the reviewer for this comment.

Regarding the circulant operators extension, the revised manuscript includes a detailed derivation in Appendix D.6. To be precise, the extension covers operators of the form x↦Ax+βx\mapsto Ax +\beta where A has equals column sums (in particular, this holds for circulant matrices), and β∈R+d\beta\in\mathbb{R}^d_+ is a background noise. Poissonian background noise is notably common in astronomy applications. The revised manuscript also includes toy experiments to illustrate this extension (similar to the one illustrating PDPS in the original manuscript): the resulting algorithm consistently outperfoms the baselines in term of Wasserstein distance. We will introduce this extension in the main text as well: it is a concrete exemple of the possibilities that arise from matching the noise structure. In particular, this specialized method relies on the Poisson structure of the likelihood and generative model and does not have a direct counterpart in a Gaussian generative model. From a practitioner point of view, opting for a Poisson generative model for Poisson inverse problems can enable specialized algorithms in certain cases, while still accomodating general forward operator.

We agree that identifying the settings in which Poisson-based generative models outperform Gaussian diffusion models is an important question, and notably that the advantages compared to Gaussian methods are not necessarely reflected in the results of Table 2. First, evaluating posterior samples remains challenging, and the reconstruction metrics reported in Table 2 should primarily be viewed as summary statistics providing a partial assessment of the methods. In our toy experiments, where more reliable evaluation metrics are available, we already observe that Poisson-based methods are substantially better suited to the problem than their Gaussian counterparts.

More importantly, the metrics reported in Table 2 assess the quality of individual samples rather than the posterior distribution as a whole. They do not necessarily capture posterior calibration or the quality of uncertainty quantification, particularly in the low-count regime, where the posterior distribution is more diffuse. For instance, a method may achieve good reconstruction scores while producing samples that are overly concentrated around the posterior mean. We illustrate this distinction through a new calibration experiment on the BrainWeb/PET dataset, presented in Appendix E.4 of our revised submission. Despite achieving competitive reconstruction metrics, Gaussian diffusion DPS substantially underestimates the presence of lesions, whereas our MPDM-based methods provide more accurate lesion detection.

We therefore believe that our experiments provide concrete evidence of the potential advantages of Poisson-based generative models for posterior sampling, even though these advantages are not systematically captured by standard image reconstruction metrics. We will clarify this point in the revised manuscript and better highlight the results of the toy experiments and the BrainWeb calibration study.

We hope that these additional explanations and results adequately address the reviewer's concerns, and we would be grateful if the reviewer could reconsider their assessment in light of this evidence.

Reviewer_jMffOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces a way of doing generative modeling using Mixed Poisson Processes (MPP). Concretely, the the authors want to sample from a target distribution π\pi

X∼πX \sim \pi Z(t)∣X∼Poisson(tX)Z(t) | X \sim \text{Poisson}(t X)

For Poisson Process, the strong law of large number tells us that

Z(t)t→a.sX\frac{Z(t)}{t} \overset{\text{a.s}}{\to}X

Thus we may sample from π\pi if we can simulate the marginal process Z(t)Z(t) for large tt since we may approximately recover XX.

To train a model to sample Z(t)Z(t) they identify the mode's intensity with the posterior mean

λt(z)=Eπ[X∣Zt−=z]\lambda_t(z) = E_{\pi} [X | Z_{t_{-}} = z]

And train a neural network to approximate this intensity kθ(z,t)≈λt(z)k_{\theta}(z, t) \approx \lambda_{t}(z).

The paper then trains this neural network by noting that pointwise minimization of the Bregmann divergence for every tt satisfies

arg⁡min⁡θEX∼π[Dϕ(Eπ[X∣Zt−],kθ(z,t))]=arg⁡min⁡θEX,Z(t)Dϕ(X,kθ(Zt−,t))\arg\min_{\theta} E_{X \sim \pi} [D_{\phi}( E_{\pi}[X | Z_{t_{-}}], k_{\theta}(z, t) ) ] = \arg\min_{\theta} E_{X, Z(t)} D_{\phi}(X, k_{\theta}(Z_{t^{-}}, t))

and end up using the L2L2-loss in their experiments.

Critical strengths

None

Critical weaknesses

I have important concerns that I think put into question the novelty of the work. In particular, there seems to be substantial overlap with the following papers which the author do not appear to cite or discuss

Generator Matching (https://arxiv.org/abs/2410.20587) Posterior Mean Matching (https://arxiv.org/abs/2412.13286)

The generative model itself appears to be mostly an incremental extension of ideas in Posterior Mean Matching while the general way of fitting the model appears to be an instantiation of ideas in Generator Matching. I'll do my best to match the other paper's notation to what the authors use and point to the relevant equations in the above papers where I think there is substantial overlap.

Posterior Mean matching considers a generative model:

X∼πX \sim \pi Zi∣X∼pαi(Z∣X)Z_i \mid X \sim p_{\alpha_{i}}(Z \mid X)

For the Gamma-Poisson PMM near equation (32) this model reads

X∼πX \sim \pi

Zt∣X∼Poisson(∑j=1tαiX)Z_t | X \sim \text{Poisson}(\sum_{j=1}^{t}\alpha_{i} X)

Compared to MDPMs model

X∼πX \sim \pi

Zt∣X∼Poisson(tX)Z_{t} | X \sim \text{Poisson}(t X)

Posterior Mean Matching introduces a Γ(a,b)\Gamma(a, b) prior and constructs the posterior mean state

μt=E[X∣Zt]=a+Ztb+At\mu_t = \mathbb{E}[X | Z_t] = \frac{a + Z_t}{b + A_t}

where Zt=∑j=1tYjZ_t = \sum_{j=1}^{t} Y_{j} and At=∑j=1tαjA_t = \sum_{j=1}^{t} \alpha_j.

I believe this is more or less exactly MDPM's Example 11. In their notation if π=Γ(α,γ)\pi = \Gamma(\alpha, \gamma) then Ki,tπ(z)=α+Zt,iγ+tK_{i,t}^{\pi}(z) = \frac{\alpha + Z_{t,i}}{\gamma + t}.

PMM then trains a neural network gϕ(μs,s)g_{\phi}(\mu_s, s) to predict the parameter governing the next noisy observation. Under the poisson likelihood training objective PMM's optimal predictor is g∗(μs,s)=E[X∣μs]g^{*}(\mu_s, s) = \mathbb{E}[X \mid \mu_s]

Compare that with MDPM which directly defines the denoiser

Ktπ(z)=E[X∣Zt−=z]K_t^{\pi}(z) = \mathbb{E}[X \mid Z_{t_{-}}=z]

and proves that this is the mean intensity of its MPP. It then trains a neural network to approximate KtπK_{t}^{\pi}. Then simulates the counting process using the learned intensity.

But recall that PMM's state is μt=a+Ztb+At  ⟺  Zt=(b+At)μt−a\mu_{t} = \frac{a + Z_t}{b + A_t} \iff Z_{t} = (b + A_t) \mu_{t} - a thus

g∗(μt,t)=KAtπ(Zt)g^{*}(\mu_{t} , t) = K_{A_{t}}^{\pi}(Z_{t})

Moreover, note that theorem 11 of PMM applied to the Gamma Poisson model reads

μt=a+Ztb+At→ZtAt→X\mu_{t} = \frac{a + Z_{t} }{b + A_{t} } \to \frac{Z_{t}}{A_{t}} \to X

With At=tA_{t} = t this is basically Theorem 2 of MDPM.

Moreover, MDPM's generator, conditional on X=xX =x has generator

Lxf(z)=∑ixi[f(z+e^i)−f(z)]\mathcal L^{x} f (z) = \sum_{i} x_{i} [f(z + \hat{e}_{i}) - f(z)]

Generator matching tells us to obtain the marginal generator by conditioning on its current state

Lf(z)=E[LXf(z)∣Zt=z]\mathcal L f(z) = \mathbb{E}[\mathcal L^{X} f(z) | Z_t = z]

substituting the above gets us

Lf(z)=∑iE[X∣Zt=z]i(f(z+ei)−f(z))\mathcal L f(z) = \sum_{i} \mathbb{E}[X | Z_{t} =z]_{i} (f(z + e_i) - f(z))

MDPM minimizes

Et,X,Z(t)Dϕ(X,kθ(Zt−,t))E_{t, X, Z(t)} D_{\phi}(X, k_{\theta}(Z_{t^{-}}, t))

while GM minimizes

Et,X,Z(t)Dϕ(FX(Zt),Ftθ(Zt))E_{t, X, Z(t)} D_{\phi}(F^{X}(Z_{t}), F_{t}^{\theta}(Z_{t}))

Simply set FX(z)=XF^{X}(z) = X and Ftθ(z)=kθ(z,t)F_{t}^{\theta}(z) = k_{\theta}(z, t).

Aside from the substantial omission of previous work, the way the paper is written makes it incredibly difficult to follow. The paper spends a significant amount of time introducing concepts from stochastic point processes, to present theorems about MDPM. The exposition can feel very terse and technical sometimes. While I understand these technicalities are needed for correctness, this can make it really challenging to really understand the significance or relevance of the results.

The experimental evaluation is also rather weak. The paper discusses earlier Poisson generative methods, including JUMP, and acknowledges similarities between their sampling algorithms. Yet it does not provide direct generative comparisons against them. This matters because the authors claim to introduce a new framework for Poissonian generative modeling. Comparing primarily against Gaussian diffusion methods cannot establish what is gained relative to existing Poisson-based approaches.

The image experiments are also extremely narrowly scoped and do not provide convincing evidence that the proposed method scales to modern generative modeling settings. Furthermore, the evaluation lacks comparisons against several directly relevant Poisson-based generative models, making it difficult to assess whether the proposed approach offers meaningful empirical improvements over existing methods. More broadly, I find the scale, breadth, and rigor of the experimental evaluation not commensurate with the scope of the paper's methodological claims. Given that the work is positioned as a general generative modeling framework, I would expect substantially more comprehensive empirical validation, including comparisons with closely related methods on standard image modeling datasets and benchmarks.

Questions for authors

It is unclear what the proposed framework contributes beyond the closely related work discussed above. I encourage the authors to provide a clear, technically precise comparison with both Posterior Mean Matching and Generator Matching, explicitly identifying which components of MPDM are genuinely new and which follow from existing constructions.

For the aspects that the authors consider novel relative to prior work, I would like to see a convincing justification of their significance, either through new theoretical insights, capabilities not afforded by existing methods, or empirical evidence demonstrating meaningful advantages over these approaches. In particular, simply presenting an alternative formulation of existing ideas would not, in my view, be sufficient to establish a substantive methodological contribution.

Based on my current understanding, I am inclined to recommend rejection. However, I remain open to reconsidering my assessment if the authors can clearly identify the contributions that go beyond the aforementioned prior work and provide compelling evidence that these contributions constitute meaningful theoretical or practical advances.

3 comments

We thank the reviewer for their feedback. In particular, we agree that the Poisson generative model considered in our paper is related to Posterior Mean Matching (PMM) [3] and Generator Matching (GM) [4]. We acknowledge these connections and have revised our Related Work section accordingly. However, we believe that the assessment of the paper's novelty should distinguish the underlying Poisson generative construction, which is related to previous works, from the contributions developed on top of this construction. In particular, the review's discussion on novelty primarily concerns the former, while our main contributions are rather the mathematical analysis of the resulting processes and their application to posterior sampling.

In particular, we do not claim in our submission to introduce a fundamentally new algorithmic principle for Poisson generative modeling, nor do we claim to introduce a new objective for learning Poisson denoisers. The paper includes notably a comparison with JUMP [2] and Itdpdm [1]. We will insist more on this aspect in the revised manuscript.

Contributions of our work

The contributions identified in our original manuscript are twofold:

  1. A counting-process formulation of Poisson generative models based on mixed Poisson processes, together with quantitative error guarantees.
  2. The application of Poisson generative models to posterior sampling in inverse problems, showing how diffusion-based posterior sampling methods, primarily developed for Gaussian diffusion models [5,6], can be extended to Poisson generative models. We consider this to be the main motivation of our work, and also why we opted for the Probabilistic Methods track.

1. Counting-process formulation and theoretical guarantees. Our approach takes the mixed Poisson process as its starting point. We then establish that the Poisson denoiser coincides with the stochastic intensity characterizing this counting process. Simulation algorithms follow naturally from this formulation, but they are not part of the definition of the model itself.

In particular, we consider two simulation procedures that leverage the learned intensity/denoiser: thinning and time discretization. Thinning allows to simulate the process without introducing additional discretization error, provided that a valid intensity envelope is available. To the best of our knowledge, [1,2,3] do not propose a similar thinning algorithm to simulate the process from the learned denoiser. [2] uses binomial thinning in its forward process, but its reverse-generation procedure is based on a discrete-time chain. The second procedure, time discretization, leads to a sampling procedure closely related to that of [2], as already acknowledged in our manuscript. Time discretization has the important practical advantage of having a fixed execution time and is perhaps more intuitive, which is why we chose it for our subsequent posterior sampling algorithms. Still, we think that thinning illustrates well the possibilities offered by the counting process formulation.

We would say that framing Poisson generative models in terms of mixed Poisson processes is analogous to studying Gaussian generative models through the theory of diffusion processes. This is certainly not the only possible approach, but it provides access to tools from stochastic analysis that have proved particularly fruitful in the Gaussian setting.

In our case, the counting process formulation enables us to derive quantitative error bounds for early stopping, denoiser estimation, and time discretization. To the best of our knowledge, related works [1,2,3] do not provide comparable quantitative bounds by controlling these different types of errors. Importantly, the proofs of several of these results rely explicitly on the counting process formulation and do not follow immediately from the similar parameterizations or sampling algorithms considered in these earlier works.

2. Posterior sampling in inverse problems.

The aforementioned works on Poisson generative modeling [1,2,3] primarily focus on unconditional generation and do not develop the posterior sampling methodology investigated in our paper.

In Section 3, we derive theoretical results and propose algorithms that use Poisson generative models as learned priors for Bayesian posterior sampling in inverse problems. These methods are evaluated experimentally in Section 4. In particular, Theorem 3 establishes an exact and natural connection between Poisson generative models and posterior sampling for simple Poisson inverse problems. For more general inverse problems, we adapt methods from the Gaussian diffusion posterior sampling literature.

This extension to posterior sampling constitutes our main algorithmic contribution. We also believe that it provides a natural and compelling motivation for studying Poisson generative models, particularly in applications where the observation noise is itself Poissonian.

Connections with Posterior Mean Matching and Generator Matching

Although developing a new Poisson generative model is not the main objective of our paper, we would like to clarify how the construction we study relates to the two references raised by the reviewer.

Connection with Posterior Mean Matching [3].

From an algorithmic perspective, we already acknowledge in our manuscript that our approach is closely related to JUMP [2], up to differences in training choices, parameterization, and simulation procedures. We believe that the same observation applies to the Gamma-Poisson specialization of PMM [3].

Indeed, to the best of our understanding, the Gamma-Poisson PMM model and JUMP rely on essentially the same underlying Poisson generative construction and closely related sampling procedures, despite differences in parameterization, training objective, and terminal estimator. In particular, the Gamma posterior mean used as the state variable in PMM is an invertible transformation of the cumulative Poisson count used in JUMP (and hence, under the corresponding time parameterization, of the mixed Poisson process state ZtZ_t, as the reviewer points out); and the trained denoisers are also directly related. It is also perhaps worth noting that JUMP [2] predates PMM [3] by approximately one year. Together, these observations support our view that JUMP is the more direct methodological antecedent to the specific Poisson generative construction considered here. This does not diminish the contribution of PMM: it introduces a more general framework for generative modeling based on posterior mean matching, encompassing several observation models. Also, we agree that PMM Theorem A.4 shows a limiting result that closely relates to our Theorem 1 (Zt/t→XZ_t/t \rightarrow X). However, our theoretical guarantees are not limited to this result, as our work also develops separate quantitative bounds controlling early stopping, denoiser estimation, and time discretization.

Connection with Generator Matching [4].

Regarding Generator Matching, the mixed Poisson process considered in our work is a Markov process characterized by the Poisson denoiser, as established in our Theorem 2. The Poisson denoiser is also directly related to the generator of the MPP, which follows from the Kolmogorov equation; we have added this observation to the Related Work paragraph. Consequently, the general framework of [4] can indeed be applied to recover the ideal generator associated with our construction; the training objectives coincide. This is an interesting connection.

However, our approach proceeds differently at the theoretical level. Rather than starting from an abstract generator-matching framework and explicitly constructing a generator, we work directly with the underlying mixed Poisson process and its stochastic intensity, using counting-process techniques to characterize its dynamics and establish quantitative approximation results. We believe that this direct probabilistic approach is both natural and useful for the class of models under consideration. It provides the precise stochastic process structure needed for our subsequent analysis.

More importantly, the general generator-matching principle does not, by itself, provide the simulation and error-analysis results established in our paper. Our work develops simulation procedures (thinning and time discretization) by leveraging the counting process formulation, and derives quantitative bounds for early stopping, denoiser estimation, and time discretization. We nevertheless agree that the connection with the general framework of Generator Matching is relevant, and we have included it in the revised Related Work section.

To summarize, although the generative construction we study is related to PMM and falls within the general scope of GM, neither connection subsumes the quantitative error analysis developed in our work or the posterior sampling methodology presented in Sections 3-4.
We hope that the clarifications provided here help make these contributions and their distinction from the related generative modelling literature clear, and that they can be taken into account in the assessment of the paper. We would be happy to discuss any remaining concerns.

[1] Sagnik Bhattacharya, Abhiram Gorle, Ahsan Bilal, Connor Ding, Amit Kumar Singh Yadav, and Tsachy Weissman. Itdpdm: Information-theoretic discrete poisson diffusion model. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, editors, Advances in Neural Information Processing Systems, volume 38, Main Conference, pages 10860–10915. Curran Associates, Inc., 2025.

[2] Tianqi Chen and Mingyuan Zhou. Learning to Jump: Thinning and Thickening Latent Counts for Generative Modeling. In Proceedings of the 40th International Conference on Machine Learning, pages 5367–5382. PMLR, July 2023. ISSN: 2640-3498.

The exposition can feel very terse and technical sometimes.

We thank the reviewer for highlighting the difficulty of following the mathematical construction of MPDMs.

We agree that the formal construction of the generative process is relatively technical and that the first part of the original manuscript may have placed too much emphasis on mathematical details at the expense of intuition.

We have moved technical details that are not essential to understanding the main construction to the appendix. We have also revised Section 2.1 to provide a more accessible introduction, including an illustrative figure highlighting the key components of the generative model

In particular, Section 2.1 now includes a more intuitive definition of the mixed Poisson process (Zt)(Z_t) and its intensity/denoiser through the intuitive relation E[Zt+h−Zt∣Zt=z]=hE[X∣Zt=z]+o(h),E[Z_{t+h}-Z_t|Z_t=z]=h E[X|Z_t=z]+o(h), where the equality is understood componentwise. It shows that the Poisson denoiser directly determines the jump rates of the process: each coordinate of the conditional expectation specifies the expected number of jumps in that coordinate over a small time interval, to first order in hh.

However, the rigorous counting-process formulation is not merely a matter of presentation. It is essential to the derivation of our quantitative error bounds, which rely on tools from point-process theory (in particular, Lemma 1 and Proposition 7).

We hope that the revised presentation offers a better balance between accessibility and mathematical rigor, while making the motivation and significance of our contributions clearer.

Comparing primarily against Gaussian diffusion methods cannot establish what is gained relative to existing Poisson-based approaches.

Our experiments do not include direct comparisons with previous Poisson generative models on unconditional generation benchmarks, nor do they establish that MPDMs outperform these models in terms of unconditional sample quality or scale to large scale image generation settings. Our experimental focus is not unconditional generation, but posterior sampling in inverse problems: in Section 3-4, we develop and implement posterior sampling algorithms based on MPDM, something that is not addressed by [1,2,3]. The literature on diffusion-based posterior sampling [5,6] is primarily built upon Gaussian diffusion models; we propose to adapt the spirit of this literature to Poisson generative models, and therefore take Gaussian diffusion-based methods as a baseline.

A direct comparison with previous Poisson models on unconditional generative modeling benchmarks could constitute an interesting complementary study, but it would not directly evaluate the posterior sampling contribution that is the focus of our experiments.

[3] Sebastian Salazar, Michal Kucer, Yixin Wang, Emily Casleton, and David Blei. Posterior mean matching: Generative modeling through online bayesian inference, 2024.

[4] Peter Holderrieth, Marton Havasi, Jason Yim, Neta Shaul, Itai Gat, Tommi Jaakkola, Brian Karrer, Ricky T. Q. Chen, and Yaron Lipman. Generator matching: Generative modeling with arbitrary markov processes. In Y. Yue, A. Garg, N. Peng, F. Sha, and R. Yu, editors, International Conference on Learning Representations, volume 2025, pages 52153–52219, 2025.

[5] Hyungjin Chung, Jeongsol Kim, Michael Thompson Mccann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. In The Eleventh International Conference on Learning Representations, 2023.

[6] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023.

Reviewer_edX3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
3
Read review
Paper summary

This paper proposes a novel diffusion generative framework based on the mixed Poisson process. The proposal has a notable benefit in improving Poisson denoising, and is shown to accommodate more general inverse problems as well.

Critical strengths

The paper is well motivated by the inverse problem applications and builds up a theory-grounded generative construction that allows tractable intensity learning and Poissonian reverse-time generation. The paper also establishes good suitability of the proposed framework to the relevant inverse problem applications. Experiment settings align with the earlier motivations.

Critical weaknesses

The current paper adopts a rigorous logic in introducing the framework, with significant efforts defining the proper stochastic process language and prioritized building up the theory first, and lack of efforts in explaining and highlighting necessary intuitions, practical implications, and the central construction. My biggest concern is that it significantly increases the burden of reading for the majority audience of the conference. I recommend a substantial rewriting to postpone the formal notations but prioritize and facilitate the understanding, by learning from how the Gaussian diffusion community writes about the framework without going through every detail of the stochastic processes. I am suggesting a weak rejection because this seems nontrivial amount of work.

Questions for authors
  • While I have no issue with the mixed reported numerical performance on the deblurring and PET tasks, I am curious if the authors have any empirical experience to share: what are the most dominant factors that explain the performance differences and whether they are amendable (or has been amended).

  • I want to know a bit more about the numerical realization of the proposed framework. For example, it seems that the intensity function learning part can be sensitive to extreme or unusual observed/simulated event count. Does a flexible neural representation tend to require more regularization to ensure the smoothness of the learned intensity function? Or it is not a problem at all. Also for the time discretization of the generation process, you mention a certain difficulty or error; is this a more severe problem than the Gaussian diffusion? Because for Gaussian diffusion, it seems to me that a lot of early efforts are placed in efficient and stable realization of the velocity or score learning, while for sampling the efforts are mostly put for acceleration rather than mitigating some intrinsic discretization error. I am just curious what's your impression of the difficulty in realizing the MDPM. And it may be helpful to discuss relevant takeaways in the manuscript as well.

4 comments

Thank you for your review, and your positive comments, stating that the paper is well motivated by the inverse problem applications and well supported by the experiments.

My biggest concern is that it significantly increases the burden of reading for the majority audience of the conference.

Thank you for your input on the difficult-to-parse mathematical construction of MPDM. We agree that the formal construction of the generative process may be relatively technical, making the first half of the paper too focused on theory and not sufficiently intuitive. We have moved technical details unnecessary to the exposition to the appendix. Section 2.1 of the revised manuscript includes an illustrative figure that highlights the core aspects of the generative model. We provide a more informal definition of a mixed Poisson process (Zt)(Z_t), and notably of its intensity/denoiser: E[Zt+h−Zt∣Zt=zt]=hE[X∣Zt=zt]+o(h)E[Z_{t+h}-Z_t|Z_t=z_t] = hE[X|Z_t=z_t] + o(h) (componentwise). This shows that this denoiser is the instantaneous "jump rate" of the process, and relates to the expected number of jumps in an infinitesimal time interval. We hope that the revised version of the manuscript helps readers to gain a quick understanding of the construction, without having to delve into details.

The precise mathematical construction is still useful: it enables us to derive theoretical guarantees by leveraging tools from point process theory (Lemma 1, Proposition 7).

What are the most dominant factors that explain the performance differences and whether they are amendable (or has been amended).

Regarding the performance differences, a difficulty in comparing Poisson-based and Gaussian-based methods is that both the posterior sampling schemes and the underlying generative models are different. We use identical architectures and training budgets for both classes of models to mitigate the effects of the generative models, but they can still influence the results. This is particularly true in the low-count setting, where observations are less informative and the prior weight is more important.

Also, the metrics reported in Table 2 evaluate the quality of individual samples, not the full posterior distribution. While it is common to use these metrics in the posterior sampling literature, they do not necessarily reflect the calibration/uncertainty quantification of the posterior; an aspect that matters in particular when the posterior is "spread" (which, in our experiments, corresponds typically to the low-count setting). For instance, image reconstruction metrics can be good even if the samples are too concentrated on the mean of the distribution. Table 2 should be viewed as a simple proxy for assessing the viability of the different methods; but ultimately the "appropriate" metric depends on the problem at hand. In particular, we report a calibration evaluation for the BrainWeb dataset in Appendix E.4 that assesses whether the sampling methods correctly detect lesions; we observe notably that despite its good image reconstruction metrics, the baseline (Gaussian) diffusion DPS is less effective compared to our MPDM-based methods (it tends to underestimate the presence of lesions).

I want to know a bit more about the numerical realization of the proposed framework. For example, it seems that the intensity function learning part can be sensitive to extreme or unusual observed/simulated event count. Does a flexible neural representation tend to require more regularization to ensure the smoothness of the learned intensity function?

In our numerical experiments, we did not observe significant differences regarding the training of the Poisson denoiser/intensity when compared to a classical Gaussian denoiser; and no extra regularization was required. In practice, the trained denoiser receives the rescaled input zt/(1+t)z_t/(1+t) (with zt∼P(tx)z_t\sim\mathcal{P}(tx)), which stabilizes the values used during training. In the experiments considered here, training a Poisson denoiser was generally no more difficult than training a Gaussian one. This observation is coherent with previous works focused on generative modelling [1,2], where good generative performance on classical datasets is obtained using a Poisson denoiser parametrized with a standard Unet architecture.

[1] Sagnik Bhattacharya, Abhiram Gorle, Ahsan Bilal, Connor Ding, Amit Kumar Singh Yadav, and Tsachy Weissman. Itdpdm: Information-theoretic discrete poisson diffusion model. In D. Belgrave, C. Zhang, H. Lin, R. Pascanu, P. Koniusz, M. Ghassemi, and N. Chen, editors, Advances in Neural Information Processing Systems, volume 38, Main Conference, pages 10860–10915. Curran Associates, Inc., 2025.

[2] Tianqi Chen and Mingyuan Zhou. Learning to Jump: Thinning and Thickening Latent Counts for Generative Modeling. In Proceedings of the 40th International Conference on Machine Learning, pages 5367–5382. PMLR, July 2023. ISSN: 2640-3498.

Also for the time discretization of the generation process, you mention a certain difficulty or error; is this a more severe problem than the Gaussian diffusion?

There are two algorithms for sampling a mixed Poisson process: thinning and time-discretization. Thinning does not induce a discretization error, but scales poorly with the dimension; we think it is worth noting that such a simulation scheme exists for MPDM, but in practice time-discretization is the most reasonable choice. Regarding time-discretization, the discretization error bound we derive in Prop 7 is of order O(Δt)O(\Delta t). Additionally, in Table 5 of the appendix, we compare the generative performance of MPDM with Gaussian diffusion on a toy distribution (with identical time discretization), and observe that the Wasserstein distance scales similarly in terms of the number of discretization steps. So, from our error bound and our experiments, we would say that the time-discretization error does not appear more severe than for Gaussian diffusion.

I am just curious what's your impression of the difficulty in realizing the MDPM. And it may be helpful to discuss relevant takeaways in the manuscript as well.

Thank you for your interest. From our experiments, our impression is that training an MPDM is generally no more difficult than training a classical Gaussian diffusion model on image datasets, at least for the relatively simple datasets considered in our study. For applications involving Poisson inverse problems, such as PET or astronomy, an MPDM can therefore be a natural choice, as its generative construction is aligned with the observation model.

The main difficulty, as the reviewer may suspect, lies in developing and analysing the methodology. This requires becoming familiar with the theory of point processes, their stochastic intensities, and the associated simulation techniques. This background may be less familiar to readers accustomed to Gaussian diffusion models, and working through it was an important part of constructing and analysing our models. However, once the mathematical construction and simulation procedures had been established and translated into explicit pseudocode, the implementation was straightforward. Implementing the proposed training and sampling algorithms does not require reproducing the underlying point process analysis. We will clarify this distinction between the theoretical construction and the practical implementation in the manuscript, and direct readers to the training and simulation pseudocode.

Thank you for your detailed and thoughtful responses and substantial work revising the manuscript. I agree with that the revised manuscript now demonstrates a better balance between technical depth and readability; though for Figure 1 I find it a bit hard to relate the caption and the depicted galaxies that jointly highlight some crucial aspects or quick intro of the proposal, which I suggest a further refinement.

I believe my other comments and questions are reasonably addressed. It is understandable that many metrics can only serve a snapshot for the overall performance, but I would appreciate it if the authors can identify some aspects worth mentioning in the main text table that help to highlight the quantitative gain of using a compatible Poissonian generative prior for the corresponding inverse problems. The experiment sections of the paper main body now looks relatively lacking necessary interpretation, which might require further compacting earlier sections for additional space.

I have also skimmed through the discussion with other reviewers, and would like to uphold the authors' perspectives in clarifying their contributions. With careful comparison, sharing relevance with the predecessor can strengthen the paper's position in deriving generative prior approaches for inverse problems. I find the paper is generally solid, provides a useful synthesis and specialized study from the MPP perspectives, and is accompanied by a suitably-chosen bounded set of experiments to demonstrates viability and usefulness. That being said, I don't actively support the following argument (though not oppose it either)

neither connection subsumes the quantitative error analysis developed in our work or the posterior sampling methodology presented in Sections 3-4.

I believe that it has become more demanding recently for the theoretical analysis to provide quantifiable algorithmic or design consequences to make it interesting or useful (a review criteria); in this regard, I find it relatively weak to extract useful values from the provided analysis alone. But my major impression is again that the paper has demonstrated sufficient informativeness and soundness as a whole, and the authors demonstrate not only technical ownership of their methodology but also insightful judgment to identify the essential tradeoffs and offer meaningful interpretations beyond the immediate questions raised.

With the above reasons, I thus increase my scores and recommend acceptance.

Thank you for the positive comments and for reevaluating your assessement.

We will adjust Figure 1 caption and also include further interpretation of the numerical experiments in Section 4.

Thank you for your engagement and continuing feedback on the paper.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revisions

We again thank all the reviewers for their valuable feedback. We would like to emphasize the different revisions made during the rebuttal period.

1. Clarity of the exposition

Section 2.1 of the original manuscript, in particular the exposition of mixed Poisson processes, was relatively formal. While the mathematical details are useful for certain theoretical results, a more intuitive exposition can help readers gain a quick understanding of the construction. We have substantively rewritten Section 2.1, postponing technical details unnecessary to the exposition to the appendix, and added a more intuitive definition of the objects. The revised manuscript also includes an illustrative figure and directs readers to pseudocode for the implementation. We believe that this version of the manuscript enables readers to grasp the important ideas without having to first delve into the mathematical details.

We also revised the Related works paragraph in Section 2.2 to be more explicit regarding the nature of our contributions compared to previous works that studied a similar generative model. We also included a comparison with additional related works that were not mentioned in the original version.

2. Algorithmic extensions and numerical experiments

Section 3-4 of the revised manuscript includes an additional posterior sampling algorithm that extends the practical implications of Theorem 3 to more general Poisson inverse problems. Additional numerical experiments illustrate this new method. Finally, we have included a calibration experiment to complement the image reconstruction metrics reported in Table 2: reconstruction metrics are a proxy for evaluating posterior sampling but remain a partial evaluation, which notably does not assess calibration. We believe these additions further illustrate the benefits of the structural match between the generative model and the inverse problem.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.