REVIEW 3 major objections 5 minor 21 references
Generative diffusion posterior sampling for informative likelihoods
T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read This paper proposes a new SMC sampler, B0SMC, for conditional diffusion sampling that constructs an observation-path twisting function and claims consistently higher effective sample size and lower sliced Wasserstein distance than DPS…
desk verdict A clean recursive Gaussian twisting for diffusion posterior SMC, but the headline efficiency claim rests on an untested zeroth-order drift approximation and a narrow synthetic experiment. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the observation path {Y_k}_{k=0}^N, initialised at the true observation and noised by the same transition p_{k|k-1} as the data, together with the time-reversed twisting $l^{{v_k}}$_k(x_k) = p_k(y_k|x_k). Its computational engine is Proposition 3.2, which iterates the operator ~K defined by replacing p_{k-1}(x_{k-1}|x_k)=N(x_{k-1}; r_k(x_k), C_k) with N(x_{k-1}; x_k, C_k); when the target likelihood is linear-Gaussian, ~K^k(f) is a Gaussian N(y_k; F_k x_k + z_k, Ω_k) with closed-form semigroup updates. This makes the guided proposal (14) tractable and gives the sampler its name, B0SMC (bridging likelihood, zeroth-order sequential Monte Carlo).
What would settle it
Compare B0SMC's effective sample size against a version of the algorithm that keeps the exact drift r_k in the recursion kernel, on a diffusion model with known closed-form r_k and step sizes large enough that r_k deviates meaningfully from identity; if ESS is similar, the identity approximation is not the source of the gains, and if ESS collapses, the approximation is the load-bearing component.
Extended reading notes
Core claim
The central claim is that the canonical twisting l^k(u_k) = ∫ f(y|u_N) q_{N|k}(u_N|u_k) du_N, which requires global approximations to the denoising transition p_0(x_0|x_k) and degrades under informative likelihoods, can be replaced by an interpolating twisting $l^{{y_k}}$_k(x_k) = p_k(y_k|x_k) that bridges from a reference likelihood at k=0 to the target likelihood at k=N. The bridge is built by noising the observation along the same forward process as the data, so that each intermediate likelihood p_k(y_k|x_k) only requires a local approximation between consecutive steps. Proposition 3.2 gives a closed-form Gaussian approximation to this twisting for linear-Gaussian likelihoods by replacing the denoising transition N(x_{k-1}; r_k(x_k), C_k) with N(x_{k-1}; x_k, C_k), yielding the recursion used in the guided proposal (14). The paper argues this keeps the SMC weights well tempered even when the observation is an outlier, and reports experiments showing B0SMC outperforms DPS, TDS, and MCGDiff on a high-dimensional Gaussian-mixture posterior.
Load-bearing premise
The load-bearing premise is that replacing each denoising step's nonlinear drift r_k by the identity map in Proposition 3.2 leaves the intermediate twisting accurate enough; if r_k is far from identity, the guided proposal is miscalibrated and the reported efficiency gains could disappear.
Editorial extensions
If this is right
- For linear-Gaussian likelihoods, the twisting sequence and guided proposal are available in closed form, so the method runs as a training-free add-on to any pre-trained diffusion model.
- Because the matrix inversions are in observation dimension c only, the per-particle cost scales independently of data dimension d, making high-dimensional posterior sampling practical.
- The method reduces the number of resampling triggers under outlier observations, implying it can reach a target effective sample size with fewer particles than TDS.
- The interpolating twisting avoids global approximations to p_0(x_0|x_k), so it should remain stable when the observation is far from the bulk of the prior, a regime where canonical twisting degenerates.
Reading between the lines
- The identity approximation in Proposition 3.2 could be upgraded to a local linearisation of r_k around x_k, preserving closed-form Gaussian updates and potentially reducing bias, but the paper does not analyse this.
- The observation-path construction suggests that the auxiliary noising schedule need not equal the data's forward process; tuning it as a control variate could further improve tempering, a test the paper does not run.
- The paper's own acknowledgement that real-image experiments are deferred means the practical claim for image restoration is unverified; a natural next experiment is a trained score network on natural images where the denoiser is strongly nonlinear.
- Because the terminal constraint is exact, B0SMC remains an asymptotically exact sampler for π(x|y) even with the approximate twisting; the approximation affects only efficiency, not the limiting target.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a new sequential Monte Carlo (SMC) sampler, termed B0SMC, for conditional sampling from a diffusion-model prior under a potentially informative or outlying Gaussian likelihood. The central idea is to introduce an auxiliary observation path {Y_k} that is noised with the same dynamics as the forward diffusion, and to construct a twisting sequence lv_k as the intermediate likelihood p_k(y_k|x_k), approximated by a closed-form Gaussian recursion. The authors show that the approximate twisting still yields an exact terminal marginal by telescoping of the Feynman-Kac product structure. The method's efficiency is compared with DPS, TDS, and MCGDiff on a 256-dimensional Gaussian-mixture OU diffusion example.
Significance. The paper addresses a relevant and recognized problem: SMC samplers for diffusion posteriors can have poor statistical efficiency when the likelihood is highly informative or the observation is an outlier. The proposed construction is conceptually interesting: instead of approximating the hard global denoising integral, it uses a local recursion that is exact for the idealized identity-drift replacement, and the terminal distribution remains exact by construction. The method is training-free, the Gaussian recursion has closed form, and the code is publicly available. These are genuine strengths. However, the central efficiency claim rests on an unvalidated zeroth-order approximation of the denoising drift, and the empirical evidence is limited to one tractable model with large reported standard deviations. If the approximation is not approximately valid for pretrained nonlinear denoisers, the claimed gains over TDS/MCGDiff may not materialize.
major comments (3)
- [Section 3.2, Proposition 3.2 and Eq. (12)-(14)] Proposition 3.2 defines the approximate operator ~K by replacing the denoising transition p_{k-1}(x_{k-1}|x_k)=N(x_{k-1}; r_k(x_k), C_k) with N(x_{k-1}; x_k, C_k), i.e., by dropping the nonlinear drift r_k. This approximation is used to compute the Gaussian twisting (13) and, through (14)-(15), the guided proposal and potentials of the B0SMC sampler. No error bound, consistency argument, or validation for nonlinear denoisers is provided. For a generic pretrained diffusion model the drift is typically not close to the identity; for example, in a DDPM-style posterior the mean is roughly (1/sqrt(alpha_k)) x_k plus a noise-prediction term, and the coefficient can be far from unity at high noise levels. Since the terminal marginal is exact by telescoping, the approximation affects only statistical efficiency, but statistical efficiency is precisely the paper's central claim. The authors should either provide a quantitative error analysis (e.g., bounds on the effect of the drift on the twisting and on the resulting weights) or empirically validate the method on a nonlinear denoiser with substantial drift, such as a VP-SDE/DDPM-type model, beyond the OU/Gaussian-mixture setting where the drift is O(Δt).
- [Equation (14)] There appears to be an index off-by-one in the guided proposal. According to the definition in Eq. (12), the denoising transition from x_k to x_{k-1} has mean r_k and covariance C_k. In the reverse-time SMC at step k, the transition is from u_{k-1}=x_{N-k+1} to u_k=x_{N-k}, so it should be described by r_{N-k+1} and C_{N-k+1}. Equation (14) instead uses r_{N-k} and C_{N-k}. This indexing error is invisible in the experiments because the test process is time-homogeneous (r and C are constant across steps), but for non-stationary schedules (e.g., a typical DDPM noise schedule) the algorithm as written uses the wrong transition. The authors should correct the indexing or clarify the time convention used in (14).
- [Section 4, Tables 1-2 and Figure 2] The empirical evidence supporting the main claim of consistent superiority is weak in its current form. In Table 1, the standard deviations are large and the intervals overlap substantially at higher outlier levels; for example, at ω=10 the SWD for B0SMC is 0.68 with standard deviation 1.83, while TDS gives 0.87 with standard deviation 1.83. No significance tests, confidence intervals, or repeated-seed summaries beyond 100 runs are provided, so the statement that B0SMC is 'consistently the best' is not statistically substantiated. Furthermore, the experiments use only a tractable Gaussian-mixture OU model, not a neural-network-trained diffusion model, despite the conclusion claiming the method applies to 'any pre-trained diffusion model'. The author's own acknowledgment that real experiments are missing supports this concern. The authors should provide significance-aware comparisons and at least one experiment with a learned nonlinear denoiser to back the general applicability claim.
minor comments (5)
- [Section 3, notation] The notation for the twisting function is confusing: the text switches between lv_{N-k}^{N-k} and lv_k^k without a clear statement of the forward-time versus reverse-time convention. A single, consistent notation with an explicit mapping between SMC step index and forward diffusion index would greatly improve readability.
- [Abstract and conclusion] The phrase 'Empirical results conclude the efficiency' is awkward and should be rephrased, and the conclusion's claim that B0SMC outperforms peer methods 'by an order of magnitude' is stronger than what Tables 1-2 show (e.g., at ω=10 the improvement over TDS is roughly 20%, not an order of magnitude).
- [Full text, line with Eq. (9)] There is a missing-space typo in the sentence 'whichstandsforanintermediatelikelihoodfunction'; it should read 'which stands for an intermediate likelihood function'.
- [Section 4, MCGDiff comparison] In the comparison with MCGDiff, the observation covariance is set to R=10^{-8} for B0SMC while MCGDiff uses a noiseless version. This is acknowledged, but the text should make explicit that this changes the effective likelihood and that the comparison is not on the identical statistical problem; a short discussion of whether this favors either method would help.
- [Appendix A and Figure 1] The figure captions and appendix would benefit from stating the resampling threshold and the number of SMC steps explicitly in the caption of Figure 1, since the ESS behavior over time is one of the paper's main supporting visualizations.
Circularity Check
No significant circularity; the twisting construction is explicit and the final marginal is exact by SMC weighting, with self-citations only contextual.
full rationale
The paper's derivation is self-contained: the twisting sequence is constructed from an auxiliary noising path and a stated recursive approximation (Propositions 3.1–3.2), not from the target posterior or from fitted parameters; the final SMC marginal is exact by the Feynman–Kac weight recursion regardless of the twisting approximation, so efficiency claims are empirical rather than definitional. The one approximation — replacing p_{k-1}(x_{k-1}|x_k)=N(x_{k-1};r_k(x_k),C_k) by N(x_{k-1};x_k,C_k) in the definition of ~K (Proposition 3.2) — is an unvalidated modeling assumption with no error bound, but it is not circular: it is an approximation to the twisting, not an input that forces the reported ESS/SWD numbers. Self-citations to Corenflos et al. (2025), Kelvinius et al. (2025), and Zhao et al. (2025) are contextual (background, exposition, inspiration) and none is used as a proof ingredient or uniqueness theorem. No equation reduces to its own input, and no fitted parameter is renamed as a prediction.
Assumptions & free parameters
free parameters (2)
- Auxiliary observation transition matrices A_k =
e^{-0.02} approximately 0.980 in the experiment; user-specified in general
- Auxiliary observation covariances Sigma_k =
approximately 0.039 in the experiment; user-specified in general
assumptions (4)
- domain assumption A generative diffusion model with Markov noising and denoising transitions q_{k|k-1} and p_{k|k-1} is available.
- domain assumption The target likelihood f(y|x) is linear Gaussian for the tractable closed-form algorithm.
- ad hoc to paper The denoising transition can be replaced by N(x_{k-1}; x_k, C_k), dropping the nonlinear denoising mean r_k.
- domain assumption Independent Brownian noises for the state and observation processes make the conditional independence structure in Proposition 3.1 valid.
invented entities (1)
-
Auxiliary observation path {Y_k}, with reverse version {V_k}
Cite this review
Pith. "Pith review of Generative diffusion posterior sampling for informative likelihoods." pith.science (2026). https://pith.science/paper/QAAUJQP2
@misc{pith2026250601083,
author = {Pith},
title = {Pith review of: Generative diffusion posterior sampling for informative likelihoods},
year = {2026},
howpublished = {\url{https://pith.science/paper/QAAUJQP2}},
note = {Machine review of arXiv:2506.01083}
}
read the original abstract
Sequential Monte Carlo (SMC) methods have recently shown successful results for conditional sampling of generative diffusion models. In this paper we propose a new diffusion posterior SMC sampler achieving improved statistical efficiencies, particularly under outlier conditions or highly informative likelihoods. The key idea is to construct an observation path that correlates with the diffusion model and to design the sampler to leverage this correlation for more efficient sampling. Empirical results conclude the efficiency.
Figures
Reference graph
Works this paper leans on
-
[1]
Albergo, M. S., Boffi, N. M., and Vanden-Eijnden, E. (2023). Stochastic interpolants: a unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797
arXiv 2023
-
[2]
Benton, J., Shi, Y., De Bortoli, V., Deligiannidis, G., and Doucet, A. (2024). From denoising diffusions to denoising M arkov models. Journal of the Royal Statistical Society Series B: Statistical Methodology , 86(2):286--301
work page 2024
-
[3]
J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., Vander P las, J., Wanderman- M ilne, S., and Zhang, Q
Bradbury, J., Frostig, R., Hawkins, P., Johnson, M. J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., Vander P las, J., Wanderman- M ilne, S., and Zhang, Q. (2018). JAX : composable transformations of P ython+ N um P y programs
2018
-
[4]
Cardoso, G., Janati, Y., Corff, S. L., and Moulines, E. (2024). Monte C arlo guided denoising diffusion models for B ayesian linear inverse problems. In Proceedings of the the 12th International Conference on Learning Representations
work page 2024
-
[5]
Chopin, N. and Papaspiliopoulos, O. (2020). An introduction to sequential M onte C arlo . Springer Series in Statistics. Springer
work page 2020
-
[6]
Chung, H., Kim, J., McCann, M. T., Klasky, M. L., and Ye, J. C. (2023). Diffusion posterior sampling for general noisy inverse problems. In Proceedings of the 11th International Conference on Learning Representations
work page 2023
-
[7]
Corenflos, A., Zhao, Z., S\" a rkk\" a , S., Sj\" o lund, J., and Sch\" o n, T. B. (2025). Conditioning diffusion models by explicit forward-backward bridging. In Proceedings of the 28th International Conference on Artificial Intelligence and Statistics , volume 258, pages 3709--3717. PMLR
work page 2025
-
[8]
Daras, G., Chung, H., Lai, C.-H., Mitsufuji, Y., Ye, J. C., Milanfar, P., Dimakis, A. G., and Delbracio, M. (2024). A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083
arXiv 2024
Show all 21 references
-
[9]
De Bortoli, V., Thornton, J., Heng, J., and Doucet, A. (2021). Diffusion S chr \"o dinger bridge with applications to score-based generative modeling. In Advances in Neural Information Processing Systems , volume 34, pages 17695--17709
2021
-
[10]
Del Moral, P. (2004). Feynman-Kac formulae: genealogical and interacting particle systems with applications . Springer New York
2004
-
[11]
and Murray, L
Del Moral, P. and Murray, L. M. (2015). Sequential monte carlo with highly informative observations. SIAM/ASA Journal on Uncertainty Quantification , 3(1):969--997
2015
-
[12]
and Song, Y
Dou, Z. and Song, Y. (2024). Diffusion posterior sampling for linear inverse problem solving: a filtering perspective. In Proceedings of the 12th International Conference on Learning Representations
2024
-
[13]
Ho, J., Jain, A., and Abbeel, P. (2020). Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems , volume 33, pages 6840--6851. Curran Associates, Inc
2020
-
[14]
Janati, Y., Moufad, B., Durmus, A., Moulines, E., and Olsson, J. (2024). Divide-and-conquer posterior sampling for denoising diffusion priors. In Advances in Neural Information Processing Systems , volume 37, pages 97408--97444. Curran Associates, Inc
2024
-
[15]
E., Zhao, Z., and Lindsten, F
Kelvinius, F. E., Zhao, Z., and Lindsten, F. (2025). Solving linear- G aussian B ayesian inverse problems with decoupled diffusion sequential M onte C arlo. In Proceedings of the 42nd International Conference on Machine Learning (ICML)
2025
-
[16]
o lund, J., and Sch\
Luo, Z., Gustafsson, F. K., Zhao, Z., Sj\" o lund, J., and Sch\" o n, T. B. (2023). Image restoration with mean-reverting stochastic differential equations. In Proceedings of the 40th International Conference on Machine Learning , volume 202, pages 23045--23066. PMLR
2023
-
[17]
o lund, J., and Sch\
Luo, Z., Gustafsson, F. K., Zhao, Z., Sj\" o lund, J., and Sch\" o n, T. B. (2025). Taming diffusion models for image restoration: a review. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences . In press
2025
-
[18]
Shen, X., Meinshausen, N., and Zhang, T. (2025). Reverse M arkov learning: Multi-step generative models for complex distributions. arXiv preprint arXiv:2502.13747
2025 arXiv
-
[19]
P., Kumar, A., Ermon, S., and Poole, B
Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. (2021). Score-based generative modeling through stochastic differential equations. In Proceedings of the 9th International Conference on Learning Representations
2021
-
[20]
L., Naesseth, C
Wu, L., Trippe, B. L., Naesseth, C. A., Blei, D., and Cunningham, J. P. (2023). Practical and asymptotically exact conditional sampling in diffusion models. In ICML 2023 Workshop on Structured Probabilistic Inference & Generative Modeling
2023
-
[21]
o lund, J., and Sch\
Zhao, Z., Luo, Z., Sj\" o lund, J., and Sch\" o n, T. B. (2025). Conditional sampling within generative diffusion models. Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences . In press
2025
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.