acceptodds
Under review as a conference paper at ICLR 2027

LatentFlow: A General Framework for Conditioning Stochastic Processes

Abstract

Stochastic-process models are, as a rule, far easier to simulate than to condition. Non-linear observations, non-Gaussian likelihoods, black-box information, and global constraints all induce intractable conditional laws, requiring bespoke, model-specific constructions. We introduce LatentFlow, a single framework for conditioning stochastic processes, with no learned neural approximations and no training. Our starting point is to write the stochastic process as the deterministic image of a tractable latent innovation, , with sampled from a simple reference distribution. This reduces process-level conditioning to latent-space inference: pull the likelihood back through , sample the resulting latent law with a tractable guided probability flow, and push the samples forward. This construction is provably exact at the level of the target law; in practice, approximation enters only through finite terminal noising, Monte Carlo guidance, and time discretisation of the continuous-time dynamics, each of which is explicit and systematically reducible. As LatentFlow is training-free, conditioning reduces to solving a single reverse-time SDE. This enables conditional sampling in seconds on a single desktop CPU across model classes that have never shared a scalable method: classical spatial priors, nonlinear stochastic dynamics, mechanistic models from the physical and life sciences, stochastic PDEs, heavy-tails and extremes, point and discrete-state processes, and neural or simulator-defined processes.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Diffusion, stochastic processes

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_uWJLOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes LatentFlow, a general method for conditional sampling from stochastic processes. A deterministic map (generator) from the latent space to the process space is used to pull back the conditioning likelihood to the latent space, and the resulting latent posterior is sampled using a guided reverse-time SDE based on a variance-preserving Gaussian noising bridge. The method is evaluated across a diverse set of stochastic-process conditioning tasks, with experiments examining its flexibility, sampling accuracy, and computational efficiency.

Critical strengths
  1. Featuring its mild latent-representation assumption and a unified construction, this method is applicable to a wide range of processes and even extends to inference over unknown model parameters.
  2. The training-free nature and relatively low computational requirements make the method promising for realistic simulation and inference scenarios.
Critical weaknesses
  1. The sampling procedure appears closely related to that of FlowGP. The main contribution therefore seems to be the extension and unification of the framework rather than a fundamentally new sampling algorithm.
  2. The scope of the claims about sampling within seconds and controllable error is not fully clear, since the practical performance depends on hyperparameters such as SS, the integration step size, and the latent dimension, as well as on the regularity and computational properties of the generator that cannot be guaranteed in all cases.
Questions for authors

Please refer to Weaknesses.

3 comments
AuthorsOpenReview ↗

Response to uWJL [Part 1/2]

Many thanks for your time in reviewing this paper, and for your helpful feedback. We are glad to hear that you found the paper interesting. We provide detailed responses to each of your comments below. In brief: in Sec. 4 we now make clearer the relationship between LatentFlow and FlowGP [1], and the specific contributions of this work. We have also made our claims regarding the computational cost and the approximation error more precise throughout the paper, and now support these with new theoretical results (App. D.3) and a comprehensive wall-clock comparison on 46 problems (App. I.5).

W1 (relation to FlowGP). We agree with the reviewer that LatentFlow is closely related to FlowGP. In particular, when the generator is affine, Tϑ(ξ)=m+LξT_{\vartheta}(\xi) = m + L\xi, and the posterior is sampled using the probability-flow ODE, LatentFlow reduces to FlowGP. We now state this explicitly in Sec. 4. We also agree that our main contribution is the extension and unification of this framework, rather than a fundamentally new sampling algorithm. We would argue, however, that this extension is still substantive:

  • LatentFlow applies to any measurable stochastic-process generator, including SDE and SPDE solvers, scale mixtures, Lévy convolutions, and black-box simulators, which may be non-invertible and non-density-evaluable (App. B.1, Prop. B.2). We now make this distinction more clearly in Sec. 4, providing a new figure (Fig. 3) which illustrates the use of LatentFlow to condition a GP (in which case it reduces to FlowGP) as well as a monotone spline process and a Lévy jump process (which cannot be conditioned using FlowGP).
  • LatentFlow allows for fully Bayesian inference over the model parameters (e.g., process, observation, and simulator parameters), jointly with the process and under the full condition, using the same sampler (Sec. 3, Prop. F.1). By contrast, FlowGP fits the kernel hyperparameters beforehand by maximum marginal likelihood, and leaves fitting them under the full condition to future work (Sec. 7 of FlowGP). In Sec. 5, we use LatentFlow to infer a kernel lengthscale and four Lotka–Volterra rates; the latter agree with MCMC references (Tab. 18, App. H.7).
  • We establish exactness of the guided SDE and ODE for a general generator TϑT_{\vartheta} (Thms. C.2–C.3), with stability bounds for the reverse SDE (Props. D.1, D.4). We also now characterise the guidance estimation error (Prop. D.8, Thm. D.11).

References

[1] Moss, H. B., et al. (2026). Conditioning Gaussian processes on almost anything. arXiv:2605.21041.

AuthorsOpenReview ↗

Response to uWJL [Part 2/2]

W2 (scope of claims). We thank the reviewer for raising this point, and agree that our original claims regarding "sampling in seconds" and "controllable error" were not sufficiently precise. We have now revised these claims, and added theoretical and empirical results which clarify how they depend on the hyperparameters, the latent dimension, and the properties of the generator.

  • Computational cost. Each draw requires K×SK \times S evaluations of the generator and its gradient, where KK is the number of integration steps and SS is the number of MC samples. The cost thus depends on the latent dimension only through the cost of a single evaluation of the generator. We have removed the claim of "sampling in seconds" from the abstract and the introduction, and Sec. 1 now states this cost instead. We also now provide a comprehensive wall-clock comparison (App. I.5), which measures the time taken by LatentFlow and a range of alternative methods to reach the "exact-sample floor" (the error attained by the same number of exact posterior samples) on 46 problems.
  • Approximation error. The error of LatentFlow decomposes into three terms: terminal initialisation, MC guidance, and time discretisation (Sec. 3, App. D). The terminal error depends only on the terminal noise level α(t1)\alpha(t_1) (Thm. D.11), and the discretisation error decreases with KK (Cor. D.5), reaching the floor for moderate KK with our default schedule and time grid (App. I.1). Meanwhile, the MC guidance error is bounded by 2GΨ/S2G\sqrt{\Psi/S} (Thm. D.11), where GG bounds the gradient of the pulled-back log-likelihood, and Ψ\Psi accumulates along the flow a quantity ϱt\varrho_t which measures how strongly the condition reweights the bridge draws. Increasing SS thus reduces this error when S≳ϱtS\gtrsim\varrho_t. Meanwhile, we can use the ESS of the guidance weights to indicate when this is not the case (App. I.2).
  • Latent dimension. The difficulty of estimating the guidance depends not on the full latent dimension, but on the number and the sharpness of the directions constrained by the condition (Cor. D.9, Prop. D.10). For example, Fig. 7 shows spatio-temporal examples with up to 122K latent dimensions, conditioned on events which constrain only a few directions. We now discuss this in Sec. 6, along with the regimes in which LatentFlow is expected to perform well (e.g., weakly informative conditions, constraints acting on few directions) or less well (e.g., many sharply constrained directions, non-affine generators), as illustrated in App. I.3.
  • Regularity of the generator. Our bounds in the process space require TϑT_{\vartheta} to be Lipschitz on the region visited by the sampler (Cor. D.7), and gradient-based guidance also requires differentiability. We now state clearly in Sec. 6 that hard constraints, discontinuous simulators, and non-smooth generators may require relaxations, surrogates, or derivative-free guidance (App. E.1), for which our theoretical guarantees are more limited.

Thank you for the detailed response. I believe the reply and the revisions have addressed my concerns, and I have raised my score from 3 to 4.

Reviewer_p55nOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors proposed a training-free framework for conditioning stochastic processes. It represents a process as a deterministic simulator driven by Gaussian random seeds, then sample posterior distribution on the random seeds via reverse-time diffusion with Monte Carlo-guided score correction.

Critical strengths

The paper provides a rather unifying framework for training-free conditioning of the stochastic processes. They supplement this framework with practical guidance techniques, which might be useful in practical situations.

Critical weaknesses

The idea of conditioning latent-space conditioning is not new, neither is training-free conditioning. The hardness of the problem lie mostly in estimating the score correction, which highly depends on the complexity of the simulator and the conditioned event. Although the authors gave some characterizations on how score correction estimation errors propagate to sampling errors, but did not provide hardness/accuracy guarantees for the score correction estimation itself. In particular, even with large number of Monte Carlo samples, the sampler may still differ a lot from the reference.

Questions for authors

Thanks for your submission. Your writing is clear ,and the proposed methodology seems interesting. I have the following questions:

  1. (Idea Novelty) Several central gradients in the methodology have existed in the literature, including diffusion reversal and conditional score decomposition. Could you clarify precisely where the contribution of the work lies? A more explicit comparison would help distinguish the truly novel ideas from the integration of existing ideas.

  2. (Guidance Estimation Hardness) It appears that practical limitations lie in hardness of estimating the guidance correction accurately and efficiently. Could you roughly characterize how this hardness depends the simulator complexity and conditional event complexity? If theoretical bounds are hard to obtain, are there practical solutions in addition to pCN corrections?

4 comments
AuthorsOpenReview ↗

Response to p55n [Part 1/3]

Many thanks for your positive feedback, as well as your constructive comments. We provide a detailed point-by-point response to your comments below. To summarise: in Secs. 1,4 we now make much clearer the specific novelty and contributions of this work. In App D.3, we now provide a bound on the approximation error for the guidance itself, as well as the resulting end-to-end sampler error. In App. I, we provide additional ablations with exact references, as well as remedies for hard cases (e.g., sharp conditioning statements).

W1 & Q1 (novelty). We agree with the reviewer that several ingredients underlying LatentFlow already exist in the literature, including innovation representations (or "non-centred parametrisations") [1], reverse-time diffusions [2], the decomposition of a conditional score into a prior score and a guidance term [2], and inference over a generator's input noise [3–10]. We now make this even more explicit in Secs. 1 and 4. Our contributions are, nonetheless, still significant. These can be summarised as follows:

  • We provide a general unifying formulation which shows how these ingredients can be applied to sample from any stochastic-process generator, including non-invertible and non-density-evaluable simulators (App. B.1, Prop. B.2), crucially without the need to learn a neural network.
  • We establish exactness of our framework for a broad class of generators TϑT_{\vartheta} (Thms. C.2–C.3), and characterise the approximation errors incurred when this framework is applied in practice (Apps D.1-D.2). We also now provide an analysis of when MC guidance is accurate (App D.3).
  • We show how to perform fully Bayesian inference for the model parameters, jointly with the process, using the same sampler (Sec. 3, Prop. F.1).

A table summarising relevant existing works, and in particular highlighting the additional novelty of LatentFlow in comparison to these works, is provided below.

IngredientExisting workLatentFlow
PriorFlowGP [11]: GP priors. Latent-noise inference [3–10]: learned neural generators.Any measurable generator: SDE and SPDE solvers, scale mixtures, Lévy convolutions, black-box simulators.
Prior scoreDiffusion guidance (DPS [12], ΠGDM [13], LGD-MC [14]): learned score, approximate posterior. <br> FlowGP: exact, but only GP generators.Exact for every generator.
GuidanceDiffusion guidance: bias does not vanish with more samples. FlowGP: MC estimate of affine map, ODE sampler.MC estimate through non-linear simulators, with explicit error (Prop. D.8, Thm. D.11); reverse SDE (Props. D.1, D.4).
Model parametersFlowGP: kernel hyperparameters fitted beforehand by maximum marginal likelihood.Sampled jointly with the process under the full condition (Prop. F.1, Sec. 5).

References

[1] Papaspiliopoulos, O., Roberts, G. O., & Sköld, M. (2007). A general framework for the parametrization of hierarchical models. Statistical Science.

[2] Song, Y., et al. (2021). Score-based generative modeling through stochastic differential equations. ICLR.

[3] Holden, M., Pereyra, M., & Zygalakis, K. C. (2022). Bayesian imaging with data-driven priors encoded by neural networks: theory, methods, and algorithms. SIAM J. Imaging Sci.

[4] Patel, D. V., Ray, D., & Oberai, A. A. (2022). Solution of physics-based Bayesian inverse problems with deep generative priors. CMAME.

[5] Purohit, V., et al. (2025). Consistency posterior sampling for diverse image synthesis. CVPR.

[6] Xia, Y., et al. (2026). Noise-adaptive diffusion sampling for inverse problems without task-specific tuning. ICLR.

[7] Achituve, I., et al. (2025). Inverse problem sampling in latent space using sequential Monte Carlo. ICML.

[8] Askari, H., et al. (2025). Latent refinement via flow matching for training-free linear inverse problem solving. NeurIPS.

[9] Whang, J., Lindgren, E. M., & Dimakis, A. G. (2021). Composing normalizing flows for inverse problems. ICML.

[10] Venkatraman, S., et al. (2025). Outsourced diffusion sampling: efficient posterior inference in latent spaces of generative models. ICML.

[11] Moss, H. B., et al. (2026). Conditioning Gaussian processes on almost anything. arXiv:2605.21041.

[12] Chung, H., et al. (2023). Diffusion posterior sampling for general noisy inverse problems. ICLR.

[13] Song, J., et al. (2023). Pseudoinverse-guided diffusion models for inverse problems. ICLR.

[14] Song, J., et al. (2023). Loss-guided diffusion models for plug-and-play controllable generation. ICML.

AuthorsOpenReview ↗

Response to p55n [Part 2/3]

W1 & Q1 (novelty), continued. To expand on this table, the reviewer is correct that the idea of latent-space conditioning is not new, as indeed we acknowledge in our related works section. The focus of existing papers, however, is on learned neural generators, whilst we show that the latent innovation representation can be used to perform sampling across a broad class of genuine stochastic process priors (although learned neural generators are indeed still covered by our framework). These works also rely on quite different sampling schemes. Several run MCMC in the noise space, using Langevin dynamics [5], HMC [4, 6], or pCN [3]; others use SMC [7], a guided flow with an approximate denoising posterior [8], or train an additional network to approximate the posterior [9, 10]. MCMC is asymptotically exact, but its accuracy after a finite number of iterations depends on the mixing time of the chain, which is generally unknown and can be very long, e.g., for multimodal or sharply constrained posteriors. LatentFlow, by contrast, is training-free and transports the prior to the posterior over a finite time horizon: each draw costs a fixed number of generator evaluations, the draws are independent, and the resulting error is bounded non-asymptotically (Thm. D.11, Cor. D.5).

A similar training-free, flow-based sampling scheme appears in FlowGP [11]. However, this work only considers GP priors. This is the special case of our framework when the generator is affine Tϑ(ξ)=m+LξT_{\vartheta}(\xi) = m + L\xi, and the posterior is sampled using the probability-flow ODE. Beyond this, our method also applies to a much broader class of stochastic process priors, including other standard spatial process priors (e.g., Student-tt, Cauchy), temporal processes (e.g., stochastic differential equations, dynamical systems with random initial conditions or exogenous forcing), spatio-temporal processes, and indeed learned neural generators (although these are not the focus of this work). We now make this point even more clearly in Sec. 4, where we provide a figure illustrating LatentFlow conditioning three stochastic process priors on four conditions, including a GP (in which case LatentFlow reduces to FlowGP), but also a monotone spline process and a Lévy jump process (in which case FlowGP cannot be applied).

Alongside revisions to Sec. 1 and Sec. 4, we have also made several further revisions to clarify the connections with these existing works. In particular, in the extended related work discussion (App. G), we now provide a more explicit comparison with training-free diffusion guidance. Further, in App I.5, we now compare (empirically) LatentFlow with the DPS [12] point estimate in place of our guidance term.

References: see Part 1/3.

AuthorsOpenReview ↗

Response to p55n [Part 3/3]

W2 & Q2 (guidance estimation). We agree with the reviewer that the difficulty of estimating the guidance is a very important point, and thank them for raising it. We have now obtained additional theoretical results which bound the error of the guidance estimator, along with the corresponding end-to-end sampling error. We also provide remedies when guidance estimation is particularly hard (e.g., for very sharp conditions).

To provide more detail on these additions, recall the following notation. For noise level tt and noised latent uu, we write qt,u=N(α(t)u,σ(t)2I)q_{t,u}=\mathcal N(\alpha(t)u,\sigma(t)^2I) for the exact denoising posterior of the innovation. We then write πt,u∝Λ qt,u\pi_{t,u}\propto\Lambda\,q_{t,u} for the tilted bridge, i.e., the bridge reweighted by the pulled-back likelihood Λ=LC∘Tϑ\Lambda=L_{\mathcal C}\circ T_\vartheta. Finally, we write g=α(t) Eπt,u[∇log⁡Λ]g=\alpha(t)\,\mathbb E_{\pi_{t,u}}[\nabla\log\Lambda] for the guidance, which is estimated from SS bridge draws.

  • In Prop D.8, we show that if ∥∇log⁡Λ∥≤G\|\nabla\log\Lambda\|\le G, then the MSE of the guidance approximation is at most 16α(t)2G2ϱt/S16\alpha(t)^2G^2\varrho_t/S, where ϱt=Eqt,u[Λ2]/Eqt,u[Λ]2=1+χ2(πt,u∥qt,u)\varrho_t=\mathbb E_{q_{t,u}}[\Lambda^2]/\mathbb E_{q_{t,u}}[\Lambda]^2=1+\chi^2(\pi_{t,u}\Vert q_{t,u}) measures how strongly the likelihood tilts the bridge. We also make the observaton that the ESS of the guidance weights ≈max⁡(1,S/ϱt)\approx\max(1,S/\varrho_t), and can be used to monitor the quantity ϱt\varrho_t for essentially no cost. We verify this numerically in App I.2.
  • In Cor D.9, we specialise to Gaussian observations of an affine generator (or, for non-linear generators, their linearisation at the bridge mean). In this case, we establish that log⁡ϱt≈12∑i:κi≫1log⁡(κi/2)\log\varrho_t\approx\frac12\sum_{i:\kappa_i\gg1}\log(\kappa_i/2) with κi=σ(t)2si2\kappa_i=\sigma(t)^2s_i^2, where sis_i are the singular values of the Jacobian of the latent-to-observation map, whitened by the observation noise. Thus, in particular, κi1/2=σ(t)si\smash{\kappa_i^{1/2}=\sigma(t)s_i} can be viewed as the spread of the bridge draws along direction ii of the observations. Thus κi≫1\kappa_i\gg1 means that the condition is much narrower than the bridge in that direction, and only a fraction of approximately κi−1/2\kappa_i^{-1/2} of the draws lands on it. The log of the required number of samples SS thus grows linearly in the number of directions that the event constrains more tightly than the bridge spread, and logarithmically in their sharpness. This means, in particular, that conditions acting on a few conditions (e.g., endpoints and thresholds) remain tractable even when they are sharp, but many simultaneous constraints are harder to estimate.
  • In Thm D.11, we show that TV≤α(t1)2σ(t1)W2(ρ,N(0,I))+2GΨ/S\mathrm{TV}\le\frac{\alpha(t_1)}{2\sigma(t_1)}W_2(\rho,\mathcal N(0,I))+2G\sqrt{\Psi/S}, where ρ\rho is the latent posterior, the first term is the terminal-noising error, and Ψ=∫β(t)α(t)2 E[ϱt] dt\Psi=\int\beta(t)\alpha(t)^2\,\mathbb E[\varrho_t]\,\mathrm dt accumulates the quantity ϱt\varrho_t (defined above) along the flow.
  • In App E.2, I.4, we propose (App. E.2) and evaluate empirically (App I.4) practical solutions when estimating the guidance is hard. One remedy is importance-sampling guidance (LF-IS): a Gaussian importance proposal fitted to the tilted bridge using Gauss-Newton. This replaces ϱ(π∥q)\varrho(\pi\Vert q) in the above results by ϱ\varrho relative to the proposal, and is exact for Gaussian conditions on affine generators. Another, as mentioned by the reviewer, is to use pCN [15] correction steps after each reverse-time diffusion step, similar in spirit to the predictor-corrector steps used in classical diffusion models [2]. In difficult cases, the combination of these approaches significantly improves sampling quality (Tab. 19, App I.4). We also discuss several other alternatives (last paragraph of App. I.4), including twisted sequential Monte Carlo (SMC) [16, 17], likelihood annealing [18, 19], and stabilised importance weights [20]. Given the already substantial length of this paper, further investigation of these approaches is left to future work.

References

[15] Cotter, S. L., Roberts, G. O., Stuart, A. M., & White, D. (2013). MCMC methods for functions: modifying old algorithms to make them faster. Statistical Science.

[16] Wu, L., et al. (2023). Practical and asymptotically exact conditional sampling in diffusion models. NeurIPS.

[17] Cardoso, G., et al. (2024). Monte Carlo guided denoising diffusion models for Bayesian linear inverse problems. ICLR.

[18] Neal, R. M. (2001). Annealed importance sampling. Statistics and Computing.

[19] Del Moral, P., Doucet, A., & Jasra, A. (2006). Sequential Monte Carlo samplers. JRSS B.

[20] Vehtari, A., et al. (2024). Pareto smoothed importance sampling. JMLR.

Thank you for the detailed responses. They have answered most of my questions.

Reviewer_7oMtOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The work is motivated by the difficulty of conditioning stochastic processes: simulating a process is easy, but modelling a conditional process is challenging and requires a model-specific sampler. Following FlowGP, the work moves the conditioning from the process to the prior noise: A conditioned process is obtained by sampling ξ0\xi_0 from the standard Gaussian tilted by the conditional likelihood and mapping it through some deterministic mapping. A training-free diffusion-styled approach is taken by the work to sample the tilted: It assumes a forward process that diffuses the conditional (tilted) distribution ξ0\xi_0 into a standard Gaussian. At a noisy state of ξ0\xi_0, to run the diffusion process in reverse, S candidate samples of untitled ξ0\xi_0 are drawn from the Gaussian conditional of the clean seed given that state, the process generated from each candidate is scored by the condition’s pull-back likelihood, the log-likelihood is back-propagated through the simulator to the seed, and the gradients are averaged with weights proportional to the candidates’ likelihood values. The work handles unknown model parameters in a similar fashion by augmenting the diffusion with model parameters. The work generalizes FlowGP (which uses an affine transformation) to an arbitrary generator of stochastic processes, and adds the estimation of model parameters. Empirical experiments show the generalizability of the approach across different processes from 2-D toy example to spatial processes to temporal processes, and across different condition types. The work shows the effectiveness of the proposed approach by comparing against existing methods as baselines in select settings.

Critical strengths
  1. The work proposed a learning-free approach, which does not rely on scoring networks or learned conditional generators and in theory is more generally applicable than a learning-based method. In this approach, the conditioning is reduced to solving the reverse of a diffusion process, with drift term estimated through self-normalized importance sampling and automatic differentiation.
  2. The theoretical results are correct and complete. For example, Appendix C systematically studied the exactness of the latent dynamics. The construction of the proposed LatentFlow approach is also built on solid existing basis including the change of variables theorem (Prop. 2.2), importance sampling, diffusion models, and automatic differentiation.
  3. The empirical evaluation is also broad covering different types of stochastic processes, and conditioning types.

Overall speaking, I think this is a solid work.

Critical weaknesses
  1. The idea proposed in this work is not novel. Methodology-wise, based on my understanding of FlowGP, the work is an extension of FlowGP from the affine mapping to an arbitrary generator. Motivation-wise, the idea of conditioning through the prior noise has been studied in existing works; Sec. 4 itself lists eight such works (Line 271-272.).
  2. Sampling efficiency is compared only in select settings. Wall-clock and ESS/s are reported against NUTS/VI on 2-D toy targets in Tab. 2-4. Can the work include sampling efficiency comparison for other settings using wall-clock time?
  3. Scalability of the approach is the biggest concern I have for the work. The reverse process relies on gradient estimation based on self-normalized importance sampling. How well the approach would work in high-dimensional data on complex manifolds (e.g. image or videos) or in settings where the target conditional is relatively sharp compared to the prior distribution is questionable. Even though the formulation of the approach is general, I would encourage the author to include a systematic discussion on when the method is expected to succeed and when it could face practical limitations.
Questions for authors

Apart from the concerns mentioned in critical weakness, I would like to ask what forward process, including the noise scheduler, is chosen by the work. What impacts on the proposed approach do we expect from the choice of the forward process? Can the author justify that theoretically or empirically?

4 comments
AuthorsOpenReview ↗

Response to 7oMt [Part 1/3]

Many thanks for your positive review and for your constructive feedback. We provide detailed responses to each of your comments below. In brief, the main changes are as follows. In Secs. 1,4 we now make clearer the specific contribution and novelty of LatentFlow. In App. D.3, we now provide theoretical results which bound the Monte Carlo (MC) guidance error and the end-to-end sampling error. In App. I, we now include additional ablations with exact references, remedies for hard cases (e.g. sharp conditioning statements), and a comprehensive wall-clock comparison on 46 problems.

W1 (novelty). We agree that the ingredients underlying LatentFlow (i.e., innovation representations, reverse-time diffusions, inference over a generator's input noise) exist. We now make this more clear in Secs. 1 and 4. Our main contributions are nonetheless significant, and can be summarised as follows:

  • We provide a general unifying formulation which shows how these ingredients can be applied to sample from any stochastic-process generator, including non-invertible and non-density-evaluable simulators (App. B.1, Prop. B.2), without training.
  • We establish, extending well beyond existing results in [1], exactness of our framework for a broad class of generators TϑT_{\vartheta} (Thms. C.2–C.3), and characterise the approximation errors incurred when this framework is applied in practice (App D.1-D.2). We also now provide an analysis of when MC guidance is accurate (App D.3).
  • We show how to perform fully Bayesian inference for the model parameters, jointly with the process, using the same framework (Sec. 3, Prop. F.1).

Regarding existing works, FlowGP [1] is indeed the special case of our framework when the generator is affine Tϑ(ξ)=m+LξT_{\vartheta}(\xi) = m + L\xi, and the posterior is sampled using the probability-flow ODE. Beyond this, however, our method also applies to a much broader class of stochastic process priors, including other standard spatial process priors (e.g., Student-tt, Cauchy), temporal processes (e.g., stochastic differential equations, dynamical systems with random initial conditions or exogenous forcing), spatio-temporal processes, and indeed learned neural generators (although these are not the focus of this work). We now make this point even more clearly in Sec. 4, where we provide a figure illustrating the use of LatentFlow to condition three stochastic process priors on four conditions, including a GP (in which case LatentFlow reduces to FlowGP), but also a monotone spline process and a Lévy jump process (in which case FlowGP cannot be applied).

References

[1] Moss, H. B., et al. (2026). Conditioning Gaussian processes on almost anything. arXiv:2605.21041.

AuthorsOpenReview ↗

Response to 7oMt [Part 2/3]

W1 (novelty), continued. The reviewer is also right to point out that the idea of conditioning through the prior noise has been studied in existing works, as acknowledged in our related works section. The focus of these papers, however, is on learned neural generators, whilst we show that the latent innovation representation can be used to perform sampling across a broad class of genuine stochastic process priors (although, naturally, learned neural generators are indeed still covered by our framework). These works also rely on quite different sampling schemes. Several run MCMC in the noise space, using Langevin dynamics [2], HMC [3, 4], or pCN [5]; others use SMC [6], a guided flow with an approximate denoising posterior [7], or train an additional network to approximate the posterior [8, 9]. MCMC is asymptotically exact, but its accuracy after a finite number of iterations depends on the mixing time of the chain, which is generally unknown and can be very long, e.g., for multimodal or sharply constrained posteriors. LatentFlow, by contrast, is training-free and transports the prior to the posterior over a finite time horizon: each draw costs a fixed number of generator evaluations, the draws are independent, and the resulting error is bounded non-asymptotically (Thm. D.11, Cor. D.5).

References

[2] Purohit, V., et al. (2025). Consistency posterior sampling for diverse image synthesis. CVPR.

[3] Patel, D. V., Ray, D., & Oberai, A. A. (2022). Solution of physics-based Bayesian inverse problems with deep generative priors. CMAME.

[4] Xia, Y., et al. (2026). Noise-adaptive diffusion sampling for inverse problems without task-specific tuning. ICLR.

[5] Holden, M., Pereyra, M., & Zygalakis, K. C. (2022). Bayesian imaging with data-driven priors encoded by neural networks: theory, methods, and algorithms. SIAM J. Imaging Sci.

[6] Achituve, I., et al. (2025). Inverse problem sampling in latent space using sequential Monte Carlo. ICML.

[7] Askari, H., et al. (2025). Latent refinement via flow matching for training-free linear inverse problem solving. NeurIPS.

[8] Whang, J., Lindgren, E. M., & Dimakis, A. G. (2021). Composing normalizing flows for inverse problems. ICML.

[9] Venkatraman, S., et al. (2025). Outsourced diffusion sampling: efficient posterior inference in latent spaces of generative models. ICML.

AuthorsOpenReview ↗

Response to 7oMt [Part 3/3]

W2 (sampling efficiency). We now provide a significantly expanded benchmark, comparing the sampling efficiency of LatentFlow versus alternative methods on every example in the paper for which we can obtain reliable ground-truth reference samples. The comparators include methods in the latent space (pCN [10], elliptical slice sampling [11], HMC [12], SMC [13], DPS-type guidance [14]), as well as methods in the process space (RWM [15], MALA [16], HMC [12], NUTS [17]). The expanded benchmark covers 46 problems, and measures the time taken to reach the "exact-sample floor" (the error attained by the same number of exact posterior samples). LatentFlow reaches the floor on 45 of these 46 problems, and on our five original benchmarks it does so 70–1400× faster than NUTS in the process space (for those examples where NUTS reaches the floor).

W3 (scalability). We thank the reviewer for raising this issue; we agree that scalability is an important consideration. We now provide a much more detailed and systematic discussion of when LatentFlow is expected to perform well or not (Sec. 6), supported by new theoretical results which characterise the MC guidance error (App D.3). We also provide additional ablations (App. I.2, I.3) which illustrate the effect of the dimension and sharpness of the condition on both the guidance estimation accuracy, and the end-to-end sampling accuracy. Beyond this, we also discuss and evaluate several practical extensions of LatentFlow, including alternative guidance estimators (App E.2, App I.4), pCN corrector steps in the spirit of the predictor-corrector schemes used in the classical diffusion-models literature [18] (App I.4), and their combination.

Q1 (forward process). We use as a default the variance preserving (VP) noising process [18] with β\beta linear on [10−5,10][10^{-5},10], so α(1)2=e−5\alpha(1)^2 = e^{-5}, and discretise the dynamics using Euler--Maruyama on a log-SNR grid. We now state these settings clearly (App. H).

The VP noising process is the natural choice in our setting: in the innovation space the prior is N(0,I)\mathcal N(0,I), and thus Gaussian noising keeps the prior score and the denoising posterior exact, with the guidance the only term that must be approximated. We have also now added theoretical (App D.4) and empirical results (App I.1) to clarify the effect of the choice of schedule. Theoretically, we show that this choice matters only through the time discretisation and the terminal level α(t1)\alpha(t_1), which trades the terminal error against the time spent at high noise, where guidance is hardest (Thm. D.11). Empirically, we compare linear, cosine, log-linear, and constant schedules in two simple settings (a GP and a bimodal bridge), as well as the choice of terminal noise and the choice of reverse-time dynamics (SDE or ODE). In both of these examples, our default settings are amongst the most accurate of all the configurations considered.

References

[10] Cotter, S. L., Roberts, G. O., Stuart, A. M., & White, D. (2013). MCMC methods for functions: modifying old algorithms to make them faster. Statistical Science.

[11] Murray, I., Adams, R. P., & MacKay, D. J. C. (2010). Elliptical slice sampling. AISTATS.

[12] Neal, R. M. (2011). MCMC using Hamiltonian dynamics. Handbook of Markov Chain Monte Carlo.

[13] Del Moral, P., Doucet, A., & Jasra, A. (2006). Sequential Monte Carlo samplers. JRSS B.

[14] Chung, H., et al. (2023). Diffusion posterior sampling for general noisy inverse problems. ICLR.

[15] Metropolis, N., et al. (1953). Equation of state calculations by fast computing machines. J. Chem. Phys.

[16] Roberts, G. O., & Tweedie, R. L. (1996). Exponential convergence of Langevin distributions and their discrete approximations. Bernoulli.

[17] Hoffman, M. D., & Gelman, A. (2014). The No-U-Turn sampler: adaptively setting path lengths in Hamiltonian Monte Carlo. JMLR.

[18] Song, Y., et al. (2021). Score-based generative modeling through stochastic differential equations. ICLR.

Reviewer_7oMtOpenReview ↗

Post-rebuttal Comment

I would like to thank the authors for the clarifications on novelty and contributions and appreciate the discussion on scalability limitations. Overall speaking the work is solid and I will maintain my positive rating.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We are grateful to all of the reviewers for their careful and constructive feedback. We have provided detailed responses to each of their points in the individual comments below. A revised version of the paper has been uploaded alongside this response, with all changes marked in green. We hope that our changes address the reviewers' concerns, and welcome any further questions, points of clarification, or discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.