acceptodds
Under review as a conference paper at ICLR 2027

FRITES: Provable Diffusion Posterior Sampling via Forward-Reverse Iterations of Tilted SDEs

Abstract

Score-based diffusion models have emerged as powerful generative frameworks, particularly suited for sampling from high-dimensional distributions. From a Bayesian point of view, the framework provides a potential solution to inverse problems, i.e., combining observations with prior data distributions learned by the diffusion model to obtain posterior sampling — a task of both high practical importance and significant theoretical challenge. Earlier works have achieved high generative quality, but lack evidence on provable posterior sampling. Recent work has sought theoretical guarantees, but faces a persistent dilemma: transport and particle-based methods are either restricted to linear settings or become computationally prohibitive in high dimensions, while annealing-based plug-and-play approaches stall in practice and fails to achieve posterior sampling. To tackle these issues, we introduce Forward-Reverse Iterations of Tilted SDEs (FRITES), a novel algorithm with provable convergence, scalable to high-dimensional data, adaptive to both linear and non-linear inverse problems, and easy to implement. In particular, FRITES introduces a Markov chain built on coupled forward and reverse diffusion dynamics, with a principled tilt so that the target posterior emerges as the unique stationary distribution. We establish in this paper both asymptotic and finite-time theoretical guarantees, and provide empirical evidence that FRITES achieves superior performance spanning synthetic low-dimensional benchmarks and high-dimensional imaging tasks comparing to state-of-the-art baseline methods.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
diffusion model, Bayesian inverse problems, Markov chain

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_r8njOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a new algorithm for Difusion Posterior Sampling. Diffusion Posterior Sampling refers to using Diffusion models for solving Bayesian inverse problems. Usually, diffusion posterior sampling ALGORITHM does not exactly model the true posterior. But this paper suggests that, if there are transition kernels (for Forward and Reverse Diffusion SDE of Eq. (9)) for the tilted transport process, we can sample from the true posterior by appling such transition kernels repeatedly. But in reality, kernel operations involve intractable integrals, so this paper rather repeats forward and reverse diffusion SAMPLING processes. For a toy problem (section 4.1), sampling can be performed analytically, but in a practical imaging problem (section 4.2), sampling is performed using a repeating sequence of (refining measurement consistency, forward diffusion, DDIM denoising). In both experiments, the proposed algorithms showed better quality of outputs.

Critical strengths

I believe the proofs and the theoretical claims are logically right, despite concerns of their usefulness.

Critical weaknesses

One weakness is that the aforementioned "kernel operation" cannot be employed in practical scenarios. Since integrating over y is actually intractable and the process relies on calculating only a single sample, the actual algorithm differs significantly from Equation (10). While Algorithm 1—designed for a toy problem where an analytical calculation of Lt\mathcal{L}_t is feasible—fares somewhat better, Algorithm 2 simplifies to a process of calculating the loss term (likelihood or measurement consistency) via gradient descent, adding noise, and then denoising using DDIM; such methods have already been explored in prior research and I see no real novelty here. Another major weakness is that the Algorithm 2 involves running a nested loop after the initial DPS process, which inevitably incurs a significant runtime penalty. Numerous studies have been proposed to reduce the runtime of DPS, and this method is expected to be considerably slower than those alternatives (though this was not explicitly reported in the paper).

Questions for authors
  1. Could you please report runtimes of your algorithms?
  2. Why do you think Theorems 2 and 3 are useful, even though they are using 'kernel' forms? Can you insist similar ones when we use just one sample y, rather than the 'distributions' on y?
  3. In which kinds of practical inverse problems your algorithm is useful?
Reviewer_CZtDOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes FRITES, a diffusion-prior posterior sampler built on tilted transport. It alternates a likelihood-tilted forward diffusion with the prior reverse diffusion, aiming to construct a posterior-invariant Markov chain without separately sampling the boosted posterior first. The paper presents asymptotic total-variation convergence, finite-step spectral bounds, a comparison of diffusion horizons, and a supplementary score-error analysis. Experiments include Gaussian-mixture models with analytical posteriors and FFHQ phase retrieval and super-resolution.

Critical strengths
  1. A useful iterative interpretation of tilted transport. The observation-informed forward process and prior reverse process provide a coherent ideal-kernel construction when the two processes are compatible. The endpoint joint-law identity gives a transparent route to posterior reversibility. Replacing the separate boosted-posterior sampling stage with repeated forward–reverse updates is a meaningful methodological idea and motivates further investigation of this framework.

  2. Controlled distributional validation with analytical references. The Gaussian-mixture experiments use analytical scores and posteriors, allowing evaluation of sampling quality separately from score-learning error. They cover multiple dimensions and observation regimes and report SWD means and standard deviations over 50 experiments. The lower reported mean SWD provides encouraging evidence for the approach in these controlled settings, beyond pointwise reconstruction quality alone.

Critical weaknesses
  1. Insufficient technical differentiation from related posterior samplers. A structured comparison of iteration rules, target distributions, linear/nonlinear applicability, approximation sources, and convergence assumptions is needed to distinguish FRITES from tilted transport and related alternating samplers. The broad statements that particle methods are impractical for high-dimensional imaging and annealed PnP methods stagnate also need qualification. In particular, the ideal SDE construction and the modified imaging algorithm should be positioned separately; otherwise, the precise methodological increment remains unclear.

  2. Inadequate treatment of the singular horizon and truncation parameter. Section 2.3 introduces δ\delta to avoid blow-up but subsequently identifies T⋆T^\star with T⋆−δT^\star-\delta. This obscures which results hold at a fixed admissible horizon and which concern a singular limit. The authors should distinguish the critical time from the working horizon and state how relevant bounds depend on the truncation.

  3. Incomplete theoretical applicability and an unquantified nonlinear-implementation gap. The applicability of Assumption 1 to the experimental settings is not established. Its accompanying claim that a Lipschitz forward operator and positive noise imply a uniformly bounded likelihood gradient is incorrect. The nonlinear extension introduces an L0L_0-based Euler approximation, but no posterior-bias bound is provided for this forward-kernel error. The imaging variant further replaces the reverse SDE with deterministic DDIM and omits MH. The guarantees for these distinct algorithms must be separated, and the tradeoff between longer-horizon mixing and small-time approximation accuracy requires analysis rather than an unqualified transfer of the ideal-kernel results.

  4. Limited evidence for generality, posterior accuracy, and computational advantage. The imaging evaluation covers only two tasks and principally compares against DPnP, with a partial DPS run serving as initialization. Table 3 shows clear improvements in reported means; however, image counts, repetition protocols, and standard deviations or confidence intervals are insufficiently specified. More directly relevant baselines and controlled ablations of the SDE, Euler, and DDIM variants are needed. Runtime, network/forward-model evaluation counts, and initialization costs should also be reported: sharing an initialization does not establish a matched-budget comparison.

Questions for authors

See Critical Weaknesses

Reviewer_7zD8OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes FRITES, a posterior sampling method that alternates a likelihood-guided forward diffusion process with a reverse process using a pretrained diffusion model. It presents asymptotic convergence and finite-step error bounds, and extends the method to nonlinear problems through a small-time approximation. Evaluation includes Gaussian mixture models with known posteriors and image reconstruction tasks involving phase retrieval and super-resolution.

Critical strengths

The forward–reverse construction is interesting. The paper identifies a likelihood-guided forward process that pairs with an existing reverse diffusion process. This provides a principled starting point for building a posterior sampler without retraining the diffusion model. The Gaussian mixture experiments directly evaluate sampling accuracy. They use analytical scores and known posteriors, with sliced-Wasserstein distance reported over 50 experiments per setting.

Critical weaknesses
  1. The image experiments test a modified algorithm without establishing its posterior accuracy. The implementation replaces the reverse SDE with DDIM, uses multiple gradient-descent steps, and omits the Metropolis correction. The paper does not establish that these changes preserve the target posterior. The reported PSNR, SSIM, and LPIPS improvements demonstrate better reconstructions, but do not establish accurate posterior sampling. This leaves a substantial gap between the theoretical contribution and the high-dimensional validation.
  2. Key steps supporting the theoretical guarantees are not established. In Appendix C.7, the finite-step convergence argument relies on square-integrability of the normalized transition kernel, but the paper does not show that this follows from Assumption 1. Separately, Section 3.3 claims that the Metropolis–Hastings correction restores exact detailed balance, while Appendix D computes the acceptance probability using an approximate density ratio. An approximate acceptance probability does not automatically provide an exact correction. These gaps directly affect the paper’s central claim of provable posterior sampling.
Questions for authors
  1. What conditions ensure the square-integrability required in Appendix C.7? Can this be proved under Assumption 1, or are additional assumptions needed?
  2. For super-resolution, which is linear, what is T∗T^*, and why not use the exact kernel?
  3. For nonlinear problems, what distribution does the chain you actually run (Euler step, no MH) converge to, and how does its bias depend on tt?
3 comments

We thank the review for his/her time and efforts.

Q1: Image experiments run a modified algorithm (DDIM, gradient steps, no MH). PSNR/SSIM/LPIPS show better reconstructions, not accurate posterior sampling.

We agree that PSNR/SSIM/LPIPS gains do not establish posterior sampling, and that the imaging algorithm departs from our theory. The revision addresses this in four ways.

(1) For linear measurements no approximation is needed: the forward kernel is the explicit Gaussian Ft(x,⋅)=N(mt(x),Σt)F_t(x,\cdot)=\mathcal{N}(m_t(x),\Sigma_t), diagonal in the right-singular basis of AA (Proposition 1). We now report this exact algorithm, FRITES-Exact, alongside the practical variant.

(2) On the Gaussian-mixture benchmarks, where the posterior is known in closed form, FRITES-Exact attains the lowest sliced-Wasserstein distance to the true posterior for d′=0.9dd'=0.9d in every dimension (e.g. 1.05 vs. 1.60 for the Euler implementation at d=80d=80). With the same initialization, ULA slightly outperforms FRITES-Euler when T∗T^* is large, consistent with Proposition 4, which shows that a small-tt Euler iteration is to first order a Langevin step. For d′=1d'=1, both FRITES variants outperform all baselines (Tables 1–2).

(3) On FFHQ super-resolution, \textsc{FRITES}-Exact attains slightly better PSNR, SSIM and LPIPS than the heuristic variant (FRITES Alg. 2), which already outperformed the baselines (see Table 3).

(4) Section 4.2 labels Algorithm 2 a heuristic, and Appendix E.2 relates it to FRITES-Exact step by step: it replaces the exact proximal step on the tilted likelihood by gradient steps on the loss, the tilt by the untilted likelihood, the covariance Σt\Sigma_t by 2tI2tI, and the stochastic reverse kernel by DDIM. With one gradient step of size 2t/σy22t/\sigma_y^2 and the stochastic reverse SDE, it reduces to FRITES-Euler, the first-order approximation of FRITES-Exact. The gradient steps and DDIM push the iterates toward point estimates, which explains Algorithm 2's higher pointwise scores and why our guarantees do not cover it.

Q2: The finite-step bound relies on square-integrability not shown to follow from Assumption 1, and an approximate MH ratio does not give exact detailed balance.

We agree that an approximate acceptance ratio does not give exact detailed balance. We have removed the correction (former Appendix D) together with the claim; it was not used in any of our experiments.

Q3:What conditions ensure the square-integrability required in Appendix C.7? Can this be proved under Assumption 1, or are additional assumptions needed?

It does not follow from Assumption 1, and positivity of the kernel alone does not imply a spectral gap on Rd\mathbb{R}^d. We now state what each result needs. Theorem 2 (asymptotic convergence) requires only Assumption 1. Theorem 3 (geometric rate) assumes a spectral gap explicitly, ∥Kt−Π∥≤λt<1\|K_t - \Pi\| \le \lambda_t < 1 on L2(q0)L^2(q_0) (Assumption 2). A checkable sufficient condition is a finite χ2\chi^2-information, ∬πt(x,z)2/(q0(x)qt(z)) dx dz<∞\iint \pi_t(x,z)^2/(q_0(x)q_t(z))\,dx\,dz<\infty (Einq. 11): it makes the operator Hilbert–Schmidt, and with strict positivity the Jentzsch–Perron theorem gives the gap (Lemma 4(b)). We prove Eq. 11 for Gaussian and Gaussian-mixture priors with a common covariance, including those of Section 4.1 (Proposition 5). Theorem 4 is adjusted accordingly: mixing is non-increasing in tt in general, and strictly improving under Eq. 11.

Q4:For super-resolution, which is linear, what is T∗T^*, and why not use the exact kernel?

We used T∗=0.01T*=0.01; for our operator and σy=0.05\sigma_y=0.05, Tc=0.02T_c= 0.02, so this tt is admissible. Using the Euler step for a linear problem was a choice made for uniformity with phase retrieval. The revision added the results using the exact kernel (Proposition 1). The new results are in Tab. 3.

Q5: For nonlinear problems, what distribution does the chain you actually run (Euler step, no MH) converge to, and how does its bias depend on tt?

To first order in tt, one Euler iteration is one unadjusted Langevin step of size 2t2t on the posterior (Proposition 4); under standard Poisson-equation conditions its stationary bias is O(t)O(t) for smooth test functions. When the posterior is strongly log-concave, Proposition 6 bounds the stationary bias by O(t)O(\sqrt{t}) in Wasserstein-2 distance, by coupling with the Langevin diffusion. Corollaries 1–2 are now stated to concern the exact kernel only, and Proposition 2(c) shows that for magnitude measurements no scheme pairing a data-consistency step with exact denoising can be exact at a fixed noise level. For nonlinear problems we therefore claim an approximate sampler with a characterized bias.

Reviewer_7zD8OpenReview ↗

response

Thanks for reply! The revision addresses my theoretical concerns: the MH correction has been removed and the spectral gap is now an explicit assumption. However, posterior sampling accuracy on images remains unverified. Regarding super-resolution, the rebuttal states a value of 24.74 dB for FRITES-Exact, whereas Table 3 states a value of 27.76 dB. Could the authors please reconcile these values and clarify whether FRITES-Exact used the gradient and DDIM steps listed in Appendix E.3. Base on this i would like to my score from 1 to 2.

AuthorsOpenReview ↗

response to reviewer comments

Thanks a lot for your feedback and pointing out the inconsistency. We have corrected the rebuttal comment to be consistent with the paper. It was made by an earlier error in the evaluation. We have checked the results again and verified the accuracy and reproducibility. The details listed in Appendix E.3 are correct. In the reverse process, we use gradient and DDIM to make the images sharp.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.