acceptodds
Under review as a conference paper at ICLR 2027

DEQ-Solver: A Posterior Score-based Deep Equilibrium ODE Solver for Inverse Problems in Medical Imaging

Abstract

Score-based Generative Models have emerged as the dominant generative paradigm. Owing to their impressive generative performance and high interpretability, they have also proven to be a compelling generative prior for solving inverse problems in medical imaging. Nevertheless, incorporating the posterior distribution derived from the physical imaging model into the sampling process continues to be a nontrivial issue. Furthermore, their prohibitive sampling cost remains a major challenge. To overcome the above limitations, we propose a posterior score-based Deep Equilibrium Solver (DEQ-Solver). It constructs a measurement-conditioned neural vector field that directly estimates the posterior score. Based on the implicit backward Euler formulation, the resulting reverse-time dynamics can be cast as a global fixed-point problem. By solving for the global equilibrium, we can jointly drive all temporal states along the entire reverse-time trajectory. Therefore, our method improves the efficiency of solving medical imaging inverse problems without compromising reconstruction quality. Across various medical imaging modalities, DEQ-Solver achieves superior reconstruction quality with substantially faster sampling, including a speedup over other accelerated samplers in Magnetic Particle Imaging.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical Image Reconstruction, Diffusion Moedels

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_qbMbOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a method for using a score-based prior for posterior sampling in ill-posed problems in medical imaging. They propose a way to learn the posterior score and then solve a global fixed-point problem (rather than a sequential reverse SDE problem) to sample from the posterior. They demonstrate their method on MPI, sparse-view CT, and compressed sensing MRI. They compare to diffusion-based methods, including DiffusionMBIR, DPS, and DDS, and classical iterative methods, including conjugate gradient (CG) and ADMM. They show generally stronger image quality (measured by PSNR and SSIM) and efficiency (measured by NFE) compared to these baselines.

Critical strengths

The idea of a using a global fixed-point method is relatively under-explored for diffusion posterior sampling and seems useful for accelerating sampling through parallelization.

Critical weaknesses

How to train the posterior score estimator is unclear. It appears that the method involves learning the score sθ([xt;AHy],t)s_\theta([x_t;A^Hy], t). Does this mean that paired (x0,y)(x_0, y) training data are required? If not, how is the standard score model sθ(xt)s_\theta(x_t) modified to produce take AHyA^Hy as additional input and output [x^0;Wt][\hat{x}_0;W_t]? Without understanding exactly how this score model is trained, it’s difficult for me to judge its soundness and whether it offers a substantial advantage over existing plug-and-play approaches.

The baseline comparisons are lacking. At this point, methods such as DAPS [1] and PnP-DM [2] are widely regarded as state-of-the-art. The authors should provide comparisons to these approaches, even if they don’t use a Tweedie-based approximation of the posterior mean.

References:

[1] Bingliang Zhang, Wenda Chu, Julius Berner, Chenlin Meng, Anima Anandkumar, Yang Song. “Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing.” CVPR 2025.

[2] Zihui Wu, Yu Sun, Yifan Chen, Bingliang Zhang, Yisong Yue, Katherine L. Bouman. “Principled Probabilistic Imaging using Diffusion Models as Plug-and-Play Priors.” NeurIPS 2024.

Questions for authors

None

3 comments
AuthorsOpenReview ↗

Response to Reviewer qbMb (Part 1/3)

We sincerely appreciate your insightful and professional comments. In the following responses, we will try our best to address these concerns.

Weakness 1: How to train the posterior score estimator is unclear.

**Response: ** Yes, our method uses paired image–measurement data (x₀, y) for training. For MPI and CT, measurements are generated from reference images using the corresponding forward operators and noise models. For MRI, we use acquired k-space measurements paired with their corresponding reference images. As we cannot include a formatted algorithm here, please refer to Algorithm 1 in Appendix A.2 for a detailed description of the training procedure. The network sθ does not directly output the posterior score. Instead, it jointly predicts a measurement-aware denoising estimate and a spatial weighting map:

sθ([xₜ; Aᴴy], t) = [x̂₀; Wₜ].

Here, Aᴴy maps the measurement into the image domain and is concatenated with xₜ along the channel dimension. Compared with an unconditional network sθ(xₜ, t), the input is extended to include this measurement information, and the output is adapted to predict both x̂₀ and Wₜ. The weights are constrained to be positive through a bounded sigmoid parameterization. These outputs are then used in the CG-based data-consistency update to approximate E[x₀ | xₜ, y], from which the posterior score is computed using Proposition 1.

We train the resulting neural vector field within the deep equilibrium framework. For each training pair, we sample a time point from the logarithmic time grid and construct adjacent diffusion states using the forward perturbation process. The noisier state serves as a fixed boundary condition, and Anderson acceleration is used to solve for the equilibrium state at the preceding time point. The training objective combines supervision of this equilibrium state against the corresponding diffusion target with supervision of the denoising output against the reference image x₀. Gradients are approximated using the one-step phantom gradient method. Thus, the measurement-conditioned network is trained jointly with the reconstruction system. The spatial weighting map is learned through this training objective, without requiring separate covariance labels.

AuthorsOpenReview ↗

Response to Reviewer qbMb (Part 2/3)

Weakness 2:it’s difficult for me to judge its soundness and whether it offers a substantial advantage over existing plug-and-play approaches.

Response: We are sorry that we did not clearly explain the rationale behind our method and its advantages over existing plug-and-play approaches.

Proposition 2 is derived under a Gaussian approximation to the denoising posterior and a linear measurement model with independent Gaussian noise. Within this model, the measurement-conditioned denoising mean is obtained by solving a linear system involving the weighting operator σ²Σₜ⁻¹. This derivation motivates the structure of our estimator. In our implementation, the denoising estimate and spatial weights are learned jointly with the reconstruction system. Rather than separately estimating Σₜ and specifying σ², we directly learn a spatial weighting map Wₜ that jointly accounts for denoising uncertainty and measurement noise. We interpret Wₜ as effective conditioning weights rather than a calibrated estimate of posterior precision. Its strictly positive entries ensure that AᴴA + diag(Wₜ) is positive definite, making the CG subproblem well-defined.

To better understand the advantages of our method, we conduct an ablation study comparing it with its plug-and-play variants. Specifically, we construct two variants using the covariance estimation strategies of ΠGDM (Song et al., 2023) and Convert (Peng et al., 2024). ΠGDM adopts a heuristic, spatially isotropic posterior variance motivated by a Gaussian assumption on x₀. Convert trains a variance prediction head and analytically converts the learned reverse-transition variances into spatially varying denoising posterior variances. Both variants combine their covariance estimates with a separately specified measurement-noise variance to construct the weighting operator. These comparisons evaluate our jointly trained formulation against plug-and-play alternatives within the same global equilibrium framework.

Table 1: Comparison of plug-and-play configurations and weighting schemes for 32-view CT reconstruction under different NFE budgets (mean ± SD). | Global Equilibrium | Plug and Play | Weighting Scheme | NFEs = 10 | | NFEs = 50 | | NFEs = 100 | | | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | | | | | PSNR ↑ | SSIM ↑ | PSNR ↑ | SSIM ↑ | PSNR ↑ | SSIM ↑ | | ✗ | ✓ | ΠGDM | 26.9 ± 1.05 | 0.722 ± 0.03 | 30.9 ± 1.42 | 0.816 ± 0.02 | 31.8 ± 1.42 | 0.831 ± 0.01 | | ✓ | ✓ | ΠGDM | 31.7 ± 0.92 | 0.856 ± 0.01 | 34.1 ± 1.15 | 0.886 ± 0.01 | 35.6 ± 0.90 | 0.908 ± 0.01 | | ✗ | ✓ | Convert | 26.6 ± 0.92 | 0.706 ± 0.03 | 30.7 ± 1.34 | 0.821 ± 0.02 | 31.6 ± 1.35 | 0.835 ± 0.02 | | ✓ | ✓ | Convert | 32.5 ± 0.92 | 0.865 ± 0.01 | 34.7 ± 0.99 | 0.896 ± 0.01 | 36.4 ± 0.78 | 0.918 ± 0.01 | | ✓ | ✗ | Wₜ | 38.4 ± 0.47 | 0.937 ± 0.01 | 37.9 ± 0.57 | 0.933 ± 0.01 | 38.2 ± 0.48 | 0.936 ± 0.01 |

As shown in Table 1, the global equilibrium framework substantially reduces the NFE budget needed to reach comparable reconstruction quality. With global equilibrium solving, ΠGDM achieves 31.7 dB at just 10 NFEs, nearly matching the 31.8 dB obtained by its non-DEQ counterpart using DDIM at 100 NFEs. Similarly, Convert achieves 32.5 dB at 10 NFEs, exceeding the 31.6 dB obtained with DDIM at 100 NFEs. These results demonstrate that the acceleration benefit also applies to plug-and-play variants Within the same global equilibrium framework, measurement-conditioned training provides further gains. Our full model achieves 38.4 dB at 10 NFEs, outperforming the ΠGDM and Convert variants by 6.7 and 5.9 dB, respectively. Increasing the budget to 50 or 100 NFEs yields no further PSNR improvement (37.9 and 38.2 dB, respectively). Thus, the DEQ framework enables comparable reconstruction quality with fewer NFEs, while joint measurement-conditioned training further improves the achievable quality.

AuthorsOpenReview ↗

Response to Reviewer qbMb (Part 3/3)

Weakness 3: The baseline comparisons are lacking.

Response: We sincerely appreciate your suggestion to include these additional baselines. We first compare our method with DAPS (Zhang et al., CVPR 2025) and PnP-DM (Wu et al., NeurIPS 2024), two state-of-the-art score-based inverse solvers. Given the strong performance of recent flow-based methods on inverse problems, we also include PnP-Flow (Martin et al., ICLR 2025) and Flower (Pourya et al., ICLR 2026). To further evaluate reconstruction quality under accelerated sampling, we compare against CoSIGN (Zhao et al., ECCV 2024) and VFM (Mammadov et al., ICLR 2026), which accelerate score-based and flow-based reconstruction, respectively. Here, we present the results for MPI reconstruction at an SNR of 10 dB. Please refer to the revised manuscript for additional results.

Table 2 MPI reconstruction at SNR = 10 dB (mean ± SD). | Method | NFEs | PSNR ↑ | SSIM ↑ | FPS ↑ | | --- | --- | --- | --- | --- | | DAPS | 2000 | 30.7 ± 4.37 | 0.907 ± 0.04 | 0.026 | | PnP-DM | 2070 | 30.9 ± 6.08 | 0.903 ± 0.09 | 0.044 | | PnP-Flow | 1000 | 32.0 ± 4.89 | 0.937 ± 0.06 | 0.085 | | Flower | 100 | 32.8 ± 4.83 | 0.938 ± 0.06 | 0.310 | | CoSIGN | 10 | 28.5 ± 4.51 | 0.621 ± 0.09 | 3.018 | | VFM | 10 | 30.4 ± 4.70 | 0.918 ± 0.05 | 6.589 | | DEQ-Solver (ours) | 10 | 32.9 ± 5.16 | 0.958 ± 0.03 | 6.330 |

Our DEQ-Solver achieves the highest PSNR and SSIM among the compared methods, reaching 32.9 dB and 0.958, respectively, with only 10 NFEs. In particular, it improves PSNR over DAPS and PnP-DM by 2.2 and 2.0 dB, respectively, while requiring substantially fewer NFEs. It also outperforms PnP-Flow, Flower on both reconstruction metrics. Compared with CoSIGN under the same budget of 10 NFEs, DEQ-Solver achieves an approximately 2.1× speedup (6.330 vs. 3.018 FPS) while improving PSNR by 4.4 dB. Although VFM achieves comparable reconstruction speed (6.589 FPS), its PSNR is 2.5 dB lower than that of DEQ-Solver. These results demonstrate a favorable trade-off between reconstruction quality and computational efficiency under low-SNR conditions.

Reviewer_Arm7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The work proposes a new approach, named DEQ-Solver to improve the sampling process of score-based generative models for inverse problems in imaging. It consists of a measurement-conditioned neural vector field for estimating posterior score and a deep equilibrium solver for solving the reverse-time dynamics as a global fixed-point problem with the support of parallelism. Experiments demonstrate that the method improves the efficiency of the sampling process without compromising reconstruction quality, on three difference datasets, i.e., MRI, CT and MPI.

Critical strengths

The method is interesting, which improves the sampling process of the score-based generative models in a possibly new way, i.e., casting it as a global fixed-point problem and formulating in a way where all trajectory states can be solved in parallel to reduce the memory cost. It also shows promising performance against other score-based baselines.

Critical weaknesses

Relevant literatures that target similar objectives as in this work are missing. It is not clear what are the existing and more recent efforts on improving the sampling efficiency of the score-based generative models. Although it mentions briefly that there are some previous methods that also support trajectory-level parallelism, no references were given and there were no discussion on the distinction of this approach. Because of that, the novelty and contribution of this work is hard to be assessed. Also, existing flow-based methods that aims to overcome the limitation of score-based models are also not mentioned and discussed, which should be, as they are close work.

Ablation study on the method itself is not provided. Existing provided ablation study is more on the varying of parameters, however, it will be interesting to see the ablation on the method itself. What if the measurement-conditioned vector field is replaced with something standard, and similar for the DEQ formulation. It is not clear which part plays a major role for the improvement.

The compared approaches are mainly classical score-based models. Comparison against recent score-based methods and flow-matching models are desired, to assess the performance of the proposed method in a better ground, especially those flow-matching based methods already provide fewer NFEs. It will be interesting to know if the proposed method can outperform the SOTA in this direction.

Questions for authors
  • Please provide relevant literatures in this field and discuss the differences and novelty of this work compared to existing research.
  • Please discuss the assumption of conditional independence of x_t and y, why are they independent given y is the measurement and x_t is the intermediate reconstruction of that measurement?
  • It is not clear what the spatial weighting map is, which does not seem to have been explained in the paper.
  • Results should be compared with some more recent approaches, which already outperform DPS etc. Comparison against flow-based methods are also desired to justify the necessity of the work, especially there are already one-step approaches.
  • Ablation study on the method itself (see weaknesses) is also necessary.
4 comments
AuthorsOpenReview ↗

Response to Reviewer Arm7 (Part 1/4)

We sincerely appreciate your insightful and constructive comments. We are encouraged by your recognition of our core idea of formulating the reverse-time dynamics as a global fixed-point problem that supports parallel evaluation of trajectory states. Your suggestions can also help us clarify the novelty and strengthen the experimental evaluation. We address your concerns point by point below. Please note that we respond jointly to overlapping weaknesses and questions to avoid repetition.

Weakness 1: Relevant literatures that target similar objectives as in this work are missing.

Response: Thank you for highlighting this issue. We are sorry that the main text did not sufficiently explain the connection to existing trajectory-parallel methods. Due to space limitations in the main manuscript, we included the comparison with DeqIR and the discussion of the differences between the two methods in the appendix. Specifically, DeqIR reformulates an explicit diffusion sampling chain as a joint fixed-point system, whereas DEQ-Solver constructs its trajectory equations using an implicit backward Euler discretization of a measurement-conditioned neural vector field. Thus, our contribution lies in this implicit formulation and its integration with measurement conditioning, rather than trajectory-level parallelism alone. As shown below, DEQ-Solver achieves higher PSNR and SSIM across all tested trajectory lengths, with the largest PSNR improvement of 4.7 dB at a trajectory length of 10. These results suggest that our implicit formulation, together with measurement conditioning, better preserves reconstruction quality under coarse temporal discretization.

Table 1: Comparison of DeqIR and DEQ-Solver for MPI reconstruction at an SNR of 20 dB (mean ± SD). | Method | Implicit | Global | Metric ↑ | Trajectory length = 10 | 20 | 50 | 100 | | --- | --- | --- | --- | --- | --- | --- | --- | | DeqIR | ✗ | ✓ | PSNR | 33.8 ± 4.27 | 35.2 ± 4.12 | 35.3 ± 4.17 | 35.1 ± 4.17 | | | | | SSIM | 0.873 ± 0.06 | 0.871 ± 0.05 | 0.853 ± 0.06 | 0.839 ± 0.06 | | DEQ-Solver (ours) | ✓ | ✓ | PSNR | 38.5 ± 4.97 | 38.2 ± 4.86 | 37.5 ± 4.76 | 37.9 ± 5.31 | | | | | SSIM | 0.980 ± 0.02 | 0.982 ± 0.02 | 0.974 ± 0.01 | 0.980 ± 0.03 |

AuthorsOpenReview ↗

Response to Reviewer Arm7 (Part 2/4)

Weakness 2: Ablation study on the method itself is not provided.

Response: Thank you for emphasizing the need for component-wise ablations. We agree that these experiments are essential for understanding the contribution of each component. We address this concern through a dedicated ablation study, as detailed below.

It is interesting to explore where our method’s improvements in reconstruction quality and speed stem from. We suggest that the improvement in reconstruction quality stems from the spatial weighting map Wₜ. The gain in reconstruction speed may arise from convergence to the global implicit equilibrium within only a few iterations. To examine these two factors, we conduct a detailed ablation study. The results are reported in the ablation table.

Ablation on the Spatial Weighting Map σ²Σₜ⁻¹

As discussed in Proposition 2, the key challenge in estimating the measurement-conditioned denoising mean is to determine the posterior covariance Σₜ of p(x₀ | xₜ) and the measurement-noise variance σ². These two quantities jointly define the weighting operator σ²Σₜ⁻¹. We compare two weighting strategies based on unconditional models: ΠGDM and Convert. ΠGDM adopts a heuristic, spatially isotropic posterior variance motivated by a Gaussian assumption on x₀. Convert instead trains a variance prediction head and analytically maps reverse-transition variances to spatially varying denoising posterior variances. As shown in the upper part of the ablation table, the conditionally learned spatial weighting map achieves better reconstruction performance at the same NFE budget. ΠGDM approximates the posterior covariance as isotropic under a Gaussian prior assumption on x₀. Although Convert estimates spatially varying posterior variances, it still requires a separately specified measurement-noise variance to construct the weighting operator. Our approach instead learns a unified spatial weighting map that jointly accounts for denoising uncertainty and measurement noise.

Ablation on Measurement Conditioning: s([xₜ; Aᴴy], t)

Furthermore, we evaluate the contribution of measurement conditioning to reconstruction performance using sθ without a weighting-map output and fixing σ²Σₜ⁻¹ = I. Measurement conditioning substantially improves reconstruction over unconditional models, while introducing spatial weighting provides additional gains. These gains are modest with DDIM (0.3 dB at 10 NFEs) but more pronounced with the global implicit equilibrium formulation (2.0 dB at 10 NFEs). This difference may reflect the integration of spatial weighting into deep equilibrium training, allowing the learned weights to adapt to the reconstruction dynamics.

Ablation on the Global Implicit Equilibrium

With the global implicit equilibrium formulation, even the unconditional models reach reconstruction performance within 10 NFEs comparable to that achieved by DDIM at 100 NFEs, as shown in the ablation table. Specifically, ΠGDM achieves 31.7 dB, compared with 31.8 dB using DDIM, while Convert achieves 32.5 dB, exceeding its DDIM counterpart at 31.6 dB. These results support faster convergence in reconstruction performance. Our full model achieves 38.4 dB at 10 NFEs, with no additional gains at 50 or 100 NFEs (37.9 and 38.2 dB, respectively). This early saturation may be due to deep equilibrium training, enabling strong reconstruction performance within a limited number of iterations.

Table 2: Ablation study on 32-view CT under different NFE budgets (mean ± SD). ✓ and ✗ indicate whether the component is enabled. | Global Equilibrium | Measurement Conditioning | Weighting Scheme | NFEs =10 | | NFEs =50 | | NFEs =100 | | | :---: | :---:| :---:| :---: | :---: | :---: | :---: | :---: | :---: | | | | | PSNR ↑ | SSIM ↑ | PSNR ↑ | SSIM ↑ | PSNR ↑ | SSIM ↑ | | ✗ | ✗ | ΠGDM | 26.9 ± 1.05 | 0.722 ± 0.03 | 30.9 ± 1.42 | 0.816 ± 0.02 | 31.8 ± 1.42 | 0.831 ± 0.01 | | ✗ | ✗ | Convert | 26.6 ± 0.92 | 0.706 ± 0.03 | 30.7 ± 1.34 | 0.821 ± 0.02 | 31.6 ± 1.35 | 0.835 ± 0.02 | | ✗ | ✓ | I | 31.4 ± 1.09 | 0.826 ± 0.03 | 32.1 ± 0.98 | 0.835 ± 0.01 | 32.1 ± 0.93 | 0.856 ± 0.01 | | ✗ | ✓ | Wₜ | 31.7 ± 0.95 | 0.841 ± 0.02 | 32.5 ± 0.98 | 0.848 ± 0.01 | 32.7 ± 1.00 | 0.851 ± 0.01 | | ✓ | ✗ | ΠGDM | 31.7 ± 0.92 | 0.856 ± 0.01 | 34.1 ± 1.15 | 0.886 ± 0.01 | 35.6 ± 0.90 | 0.908 ± 0.01 | | ✓ | ✗ | Convert | 32.5 ± 0.92 | 0.865 ± 0.01 | 34.7 ± 0.99 | 0.896 ± 0.01 | 36.4 ± 0.78 | 0.918 ± 0.01 | | ✓ | ✓ | I | 36.4 ± 0.78 | 0.918 ± 0.01 | 36.8 ± 0.68 | 0.921 ± 0.01 | 37.2 ± 0.61 | 0.926 ± 0.02 | | ✓ | ✓ | Wₜ | 38.4 ± 0.47 | 0.937 ± 0.01 | 37.9 ± 0.57 | 0.933 ± 0.01 | 38.2 ± 0.48 | 0.936 ± 0.01 |

AuthorsOpenReview ↗

Response to Reviewer Arm7 (Part 3/4)

***Weakness 3: The compared approaches are mainly classical score-based models. ***

Response: Thank you for this constructive suggestion. We agree that comparisons with recent score-based and flow-matching inverse solvers, particularly those designed for low-NFE reconstruction, provide a more comprehensive assessment of our method’s reconstruction quality and efficiency.

To address this concern, we expand our comparisons to include recent flow-based inverse solvers, PnP-Flow (Martin et al., ICLR 2025) and Flower (Pourya et al., ICLR 2026). We also include two advanced score-based inverse solvers, DAPS (Zhang et al., CVPR 2025) and PnP-DM (Wu et al., NeurIPS 2024). Furthermore, to evaluate reconstruction quality under accelerated sampling, we compare against CoSIGN (Zhao et al., ECCV 2024) and VFM (Mammadov et al., ICLR 2026), which represent score-based and flow-based acceleration methods, respectively. Here, we present only the results for MPI reconstruction. Please refer to the revised manuscript for additional results.

Table 3: MPI reconstruction at SNR = 10 dB (mean ± SD). | Method | NFEs | PSNR ↑ | SSIM ↑ | FPS ↑ | | --- | --- | --- | --- | --- | | DAPS | 2000 | 30.7 ± 4.37 | 0.907 ± 0.04 | 0.026 | | PnP-DM | 2070 | 30.9 ± 6.08 | 0.903 ± 0.09 | 0.044 | | PnP-Flow | 1000 | 32.0 ± 4.89 | 0.937 ± 0.06 | 0.085 | | Flower | 100 | 32.8 ± 4.83 | 0.938 ± 0.06 | 0.310 | | CoSIGN | 10 | 28.5 ± 4.51 | 0.621 ± 0.09 | 3.018 | | VFM | 10 | 30.4 ± 4.70 | 0.918 ± 0.05 | 6.589 | | DEQ-Solver (ours) | 10 | 32.9 ± 5.16 | 0.958 ± 0.03 | 6.330 |

As shown in the table, DEQ-Solver achieves the highest PSNR and SSIM for MPI reconstruction at an SNR of 10 dB, reaching 32.9 dB and 0.958, respectively, with only 10 NFEs. It outperforms methods requiring substantially more NFEs, including DAPS, PnP-DM, PnP-Flow and Flower. Compared with CoSIGN under the same budget of 10 NFEs, DEQ-Solver achieves an approximately 2.1× speedup (6.330 vs. 3.018 FPS) while improving PSNR by 4.4 dB. Although VFM achieves comparable reconstruction speed (6.589 FPS), its PSNR is 2.5 dB lower than that of DEQ-Solver. These results demonstrate a favorable trade-off between reconstruction quality and computational efficiency under low-SNR conditions.

Furthermore, we compare DEQ-Solver with CoSIGN and VFM at an SNR of 20 dB across NFE budgets ranging from 1 to 100. The results are reported in Table 4. We also provide PSNR-versus-NFE curves for MPI reconstruction in Appendix A.5 to illustrate these trends and refer the reviewer to that section for further details. CoSIGN achieves efficient sampling through consistency distillation, combined with ControlNet guidance. Its PSNR increases from 22.8 dB at 1 NFE to 35.3 dB at 10 NFEs, but decreases with larger budgets. Its training comprises diffusion teacher training, consistency distillation, and subsequent ControlNet training, introducing additional training computation. VFM achieves 31.9 dB with just 1 NFE, demonstrating its effectiveness for single-step reconstruction. However, its PSNR remains between 31.9 and 32.1 dB across the evaluated budgets. VFM incorporates measurement information into the initial noise distribution through an observation-conditioned noise adapter jointly trained with a flow map. Subsequent evaluations propagate the sample through the learned flow without explicit measurement-based corrections at each step. This may help explain the limited gains from increasing NFEs observed in our experiments. Our method achieves PSNRs of 32.2 and 32.5 dB at 1 and 5 NFEs, respectively, followed by a substantial improvement to 38.5 dB at 10 NFEs. At 1 NFE, it achieves slightly higher PSNR than VFM but lower SSIM (0.868 versus 0.934). Its improvement with additional evaluations is consistent with the iterative refinement used in its fixed-point formulation. At 10 NFEs, DEQ-Solver outperforms both baselines on PSNR and SSIM across all their evaluated budgets, including 100 NFEs. Increasing its budget to 50 and 100 NFEs yields PSNRs of 38.5 and 38.6 dB, respectively, indicating that reconstruction quality largely saturates by 10 NFEs.

Table 4: Comparison of VFM, CoSIGN, and DEQ-Solver for MPI reconstruction at an SNR of 20 dB under different NFE budgets (mean ± SD). | Method | Metric ↑ | NFEs = 1 | NFEs = 5 | NFEs = 10 | NFEs = 50 | NFEs = 100 | | --- | --- | --- | --- | --- | --- | --- | | VFM | PSNR | 31.9 ± 4.75 | 32.0 ± 4.86 | 32.1 ± 4.86 | 32.0 ± 4.86 | 31.9 ± 4.85 | | | SSIM | 0.934 ± 0.04 | 0.935 ± 0.04 | 0.935 ± 0.04 | 0.935 ± 0.04 | 0.934 ± 0.04 | | CoSIGN | PSNR | 22.8 ± 3.70 | 33.4 ± 4.32 | 35.3 ± 4.49 | 34.9 ± 5.14 | 33.5 ± 5.45 | | | SSIM | 0.571 ± 0.11 | 0.854 ± 0.07 | 0.865 ± 0.06 | 0.814 ± 0.08 | 0.785 ± 0.08 | | DEQ-Solver (ours) | PSNR | 32.2 ± 4.05 | 32.5 ± 3.91 | 38.5 ± 4.97 | 38.5 ± 5.25 | 38.6 ± 5.16 | | | SSIM | 0.868 ± 0.04 | 0.920 ± 0.03 | 0.980 ± 0.02 | 0.983 ± 0.02 | 0.979 ± 0.02 |

AuthorsOpenReview ↗

Response to Reviewer Arm7 (Part 4/4)

Question 1: Please discuss the assumption of conditional independence of x_t and y, why are they independent given y is the measurement and x_t is the intermediate reconstruction of that measurement?

Response: Thank you for this thoughtful question. Given the clean image x₀, the measurement y and the diffusion state xₜ are generated through the physical measurement process and the forward diffusion process, respectively: y = Ax₀ + n, xₜ = sqrt(ᾱₜ)x₀ + sqrt(1 − ᾱₜ)ε. The measurement noise n and diffusion noise ε are assumed to be independent of each other and of x₀. Given x₀, the measurement y varies only with the measurement noise n, while the diffusion state xₜ varies only with the independently generated diffusion noise ε. Therefore, xₜ and y are conditionally independent given x₀: p(xₜ, y | x₀) = p(xₜ | x₀)p(y | x₀). Equivalently, p(xₜ | x₀, y) = p(xₜ | x₀): once x₀ is known, additionally observing y does not change the conditional distribution of xₜ.

Although xₜ serves as an intermediate state during reconstruction, this conditional independence follows from the forward generative model used to derive the posterior score. During reconstruction, x₀ is unknown, and the measurement y informs its estimation through the conditional denoising mean E[x₀| xₜ, y], which determines the posterior score in Proposition 1.

Question 2: It is not clear what the spatial weighting map is, which does not seem to have been explained in the paper.

Response: We are sorry that the spatial weighting map was not clearly described. We explain its formulation and role below and have revised the manuscript accordingly. The spatial weighting map Wₜ is an image-sized map of positive weights that controls the relative contributions of the denoising estimate and measurement consistency at each spatial location.

Its formulation is motivated by Proposition 2. After rewriting the conditional denoising mean as a linear system, the denoising posterior covariance Σₜ and measurement-noise variance σ² enter through the weighting operator σ²Σₜ⁻¹. Under a diagonal covariance approximation:

Σₜ = diag(vₜ,₁, …, vₜ,ₕw),

σ²Σₜ⁻¹ = diag(σ²/vₜ,₁, …, σ²/vₜ,ₕw).

These diagonal entries can be arranged into an h × w spatial weighting map. Motivated by this correspondence, our network jointly predicts a measurement-aware denoising estimate x̂₀ and Wₜ:

sθ([xₜ; Aᴴy], t) = [x̂₀; Wₜ].

We learn the effective weights directly, without separately estimating σ² and Σₜ or constructing the full covariance matrix. The predicted weights are constrained to be strictly positive using a bounded sigmoid parameterization. They are then incorporated into the linear system solved approximately by CG:

[diag(Wₜ) + AᴴA] x̂₀|ₜ,y = Wₜ ⊙ x̂₀ + Aᴴy.

Here, ⊙ denotes element-wise multiplication. The notation x̂₀ refers to the network’s direct denoising output, whereas x̂₀|ₜ,y denotes the estimate after the CG-based data-consistency update. Both incorporate measurement information, since the network also receives Aᴴy. The latter estimate is used to approximate E[x₀ | xₜ, y] and thereby estimate the posterior score through Proposition 1.

Larger weights impose a stronger penalty on deviations from the denoising estimate at the corresponding locations, while smaller weights allow greater adjustment driven by measurement consistency. Thus, Wₜ functions as a learned spatially adaptive regularization weight, motivated by the operator σ²Σₜ⁻¹ that combines denoising uncertainty and measurement noise.

Reviewer_GiczOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. Strong empirical performance. The method achieves competitive or superior reconstruction quality across three medical imaging modalities, with substantial PSNR improvements in CT and MRI.
  2. An effective combination of numerical techniques. Integrating measurement-conditioned denoising, CG-based data consistency, and global equilibrium solving provides a promising approach to improving the reconstruction quality–efficiency trade-off.
Critical strengths

None

Critical weaknesses
  1. Insufficient discussion of closely related work and limited methodological novelty. The proposed posterior-score estimation framework closely resembles Improving Diffusion Models for Inverse Problems Using Optimal Posterior Covariance (Peng et al., ICML 2024). Both employ Gaussian approximations to the denoising posterior, covariance-aware conditional mean estimation, and CG-based data consistency. The prior work also investigates learned covariance representations. The authors should discuss this substantial overlap, clarify their specific methodological contributions, and ideally provide a direct comparison.
  2. Missing ablation on the measurement-conditioned denoiser. Unlike conventional DDS and other diffusion-based inverse solvers that use unconditional pretrained denoisers, the proposed method trains a measurement-conditioned network using paired data. Therefore, the reported improvements may largely originate from this stronger conditional network rather than the proposed DEQ solver. A critical missing experiment is to directly incorporate the authors' trained denoiser into conventional DDS with DDIM or DPM-Solver++ sampling. An additional comparison using the proposed spatially weighted CG correction with conventional sequential sampling would help isolate the contribution of the global DEQ formulation. Without these controlled comparisons, the source of the performance gains remains unclear.
Questions for authors
  1. DDS baseline configuration. What exactly does "DDS Baseline" (1000 NFEs) refer to in Tables 1–3? The original DDS primarily demonstrates accelerated DDIM-based sampling with substantially fewer steps and allows stochasticity to be controlled through η\eta. Please clarify the sampling algorithm, stochasticity settings, and hyperparameter tuning for each DDS variant, and explain how these configurations relate to the original implementation.
  2. Additional baselines. The comparisons would benefit from more recent diffusion-based inverse solvers, such as DAPS, DiffPIR, and DC-PnPDP, where applicable. In particular, the MRI experiments lack comparisons with representative supervised unrolled reconstruction methods such as MoDL or VarNet, which are relevant given the proposed method's measurement-conditioned training.
  3. Interpretation of the posterior score. Proposition 2 assumes that the denoising mean is E[x0∣xt]E[x_0\mid x_t], whereas the proposed network additionally takes AHyA^Hy as input. How does this affect the probabilistic justification of Eq. (12)–(16)? Furthermore, is the learned WtW_t intended to estimate the posterior precision, or does it function as a learned regularization weight?
  4. Global equilibrium solving and computational cost. The global fixed-point formulation is algebraically equivalent to sequential backward Euler when solved exactly. What explains their substantial performance difference in Table 5? Could the authors report convergence residuals and clarify the effective computational cost, given that each DEQ iteration evaluates multiple temporal states and performs CG updates?
6 comments
AuthorsOpenReview ↗

Response to Reviewer Gicz (Part 1/5)

We sincerely appreciate your insightful and professional comments. In the following responses, we will make every effort to address your concerns.

Weakness 1: Insufficient discussion of closely related work and limited methodological novelty.

Response: Thank you for highlighting this closely related work and the need to clarify our specific methodological contributions. We acknowledge that both approaches employ Gaussian approximations to the denoising posterior and CG-based data consistency.

The method (Peng et al., ICML 2024) learns the denoising posterior covariance Σₜ in the discrete wavelet transform domain. The resulting covariance captures spatially varying uncertainty and pixel correlations in the clean image conditioned on the diffusion state at time step t. However, it still requires a separately specified measurement-noise variance σ² for CG-based data consistency. In our formulation, the denoising posterior covariance and measurement-noise variance are combined into a single weighting operator, σ²Σₜ⁻¹. Under a diagonal covariance approximation, its diagonal entries form an image-sized spatial weighting map Wₜ. The distinction therefore lies in directly learning the effective conditioning weights, rather than in introducing posterior covariance learning itself. Furthermore, we couple the learned spatial weighting map with the continuous-time neural vector field, yielding a global implicit equilibrium formulation. This formulation enables more accurate reconstructions under smaller NFE budgets.

Table 1: Comparison of DWT-Var and DEQ-Solver for MPI reconstruction at different SNR levels (mean ± SD), including throughput (FPS). | Method | NFEs | FPS ↑ | SNR = 10 dB | | SNR = 20 dB | | SNR = 30 dB | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | | | PSNR ↑ | SSIM ↑ | PSNR ↑ | SSIM ↑ | PSNR ↑ | SSIM ↑ | | DWT-Var | 50 | 0.113 | 31.7 ± 4.52 | 0.931 ± 0.05 | 35.2 ± 3.78 | 0.967 ± 0.03 | 35.9 ± 3.67 | 0.971 ± 0.03 | | DEQ-Solver (ours) | 10 | 6.330 | 32.9 ± 5.16 | 0.958 ± 0.03 | 38.5 ± 4.97 | 0.980 ± 0.02 | 42.7 ± 5.22 | 0.988 ± 0.01 |

To further evaluate our method against DWT-Var, we compare their MPI reconstruction performance across different SNR levels. As shown in the table, our method achieves higher mean PSNR and SSIM across all three SNR levels with a lower NFE budget. Furthermore, our method achieves an approximately 56× speedup over DWT-Var (6.330 FPS vs. 0.113 FPS). These results demonstrate the advantages of our method and support the effectiveness of our proposed contributions.

AuthorsOpenReview ↗

Response to Reviewer Gicz (Part 2/5)

Weakness 2: Missing ablation on the measurement-conditioned denoiser.

Response: We sincerely appreciate your insightful and professional comments. We found this comment to be important and useful. This comment can improve the quality and persuasiveness of our paper. We address this concern through a dedicated ablation study, as detailed below.

It is interesting to explore where our method’s improvements in reconstruction quality and speed stem from. We suggest that the improvement in reconstruction quality stems from the spatial weighting map Wₜ. The gain in reconstruction speed may arise from convergence to the global implicit equilibrium within only a few iterations. To examine these two factors, we conduct a detailed ablation study. The results are reported in the ablation table.

Ablation on the Spatial Weighting Map σ²Σₜ⁻¹

As discussed in Proposition 2, the key challenge in estimating the measurement-conditioned denoising mean is to determine the posterior covariance Σₜ of p(x₀ | xₜ) and the measurement-noise variance σ². These two quantities jointly define the weighting operator σ²Σₜ⁻¹. We compare two weighting strategies based on unconditional models: ΠGDM and Convert. ΠGDM adopts a heuristic, spatially isotropic posterior variance motivated by a Gaussian assumption on x₀. Convert instead trains a variance prediction head and analytically maps reverse-transition variances to spatially varying denoising posterior variances. As shown in the upper part of the ablation table, the conditionally learned spatial weighting map achieves better reconstruction performance at the same NFE budget. ΠGDM approximates the posterior covariance as isotropic under a Gaussian prior assumption on x₀. Although Convert estimates spatially varying posterior variances, it still requires a separately specified measurement-noise variance to construct the weighting operator. Our approach instead learns a unified spatial weighting map that jointly accounts for denoising uncertainty and measurement noise.

Ablation on Measurement Conditioning: s([xₜ; Aᴴy], t)

Furthermore, we evaluate the contribution of measurement conditioning to reconstruction performance using sθ without a weighting-map output and fixing σ²Σₜ⁻¹ = I. Measurement conditioning substantially improves reconstruction over unconditional models, while introducing spatial weighting provides additional gains. These gains are modest with DDIM (0.3 dB at 10 NFEs) but more pronounced with the global implicit equilibrium formulation (2.0 dB at 10 NFEs). This difference may reflect the integration of spatial weighting into deep equilibrium training, allowing the learned weights to adapt to the reconstruction dynamics.

Ablation on the Global Implicit Equilibrium

With the global implicit equilibrium formulation, even the unconditional models reach reconstruction performance within 10 NFEs comparable to that achieved by DDIM at 100 NFEs, as shown in the ablation table. Specifically, ΠGDM achieves 31.7 dB, compared with 31.8 dB using DDIM, while Convert achieves 32.5 dB, exceeding its DDIM counterpart at 31.6 dB. These results support faster convergence in reconstruction performance. Our full model achieves 38.4 dB at 10 NFEs, with no additional gains at 50 or 100 NFEs (37.9 and 38.2 dB, respectively). This early saturation may be due to deep equilibrium training, enabling strong reconstruction performance within a limited number of iterations.

Table 2: Ablation study on 32-view CT under different NFE budgets (mean ± SD). ✓ and ✗ indicate whether the component is enabled. | Global Equilibrium | Measurement Conditioning | Weighting Scheme | NFEs =10 | | NFEs =50 | | NFEs =100 | | | :---: | :---: | :---: | :---: | :--- | :---: | :---: | :---: | :---: | | | | | PSNR ↑ | SSIM ↑ | PSNR ↑ | SSIM ↑ | PSNR ↑ | SSIM ↑ | | ✗ | ✗ | ΠGDM | 26.9 ± 1.05 | 0.722 ± 0.03 | 30.9 ± 1.42 | 0.816 ± 0.02 | 31.8 ± 1.42 | 0.831 ± 0.01 | | ✗ | ✗ | Convert | 26.6 ± 0.92 | 0.706 ± 0.03 | 30.7 ± 1.34 | 0.821 ± 0.02 | 31.6 ± 1.35 | 0.835 ± 0.02 | | ✗ | ✓ | I | 31.4 ± 1.09 | 0.826 ± 0.03 | 32.1 ± 0.98 | 0.835 ± 0.01 | 32.1 ± 0.93 | 0.856 ± 0.01 | | ✗ | ✓ | Wₜ | 31.7 ± 0.95 | 0.841 ± 0.02 | 32.5 ± 0.98 | 0.848 ± 0.01 | 32.7 ± 1.00 | 0.851 ± 0.01 | | ✓ | ✗ | ΠGDM | 31.7 ± 0.92 | 0.856 ± 0.01 | 34.1 ± 1.15 | 0.886 ± 0.01 | 35.6 ± 0.90 | 0.908 ± 0.01 | | ✓ | ✗ | Convert | 32.5 ± 0.92 | 0.865 ± 0.01 | 34.7 ± 0.99 | 0.896 ± 0.01 | 36.4 ± 0.78 | 0.918 ± 0.01 | | ✓ | ✓ | I | 36.4 ± 0.78 | 0.918 ± 0.01 | 36.8 ± 0.68 | 0.921 ± 0.01 | 37.2 ± 0.61 | 0.926 ± 0.02 | | ✓ | ✓ | Wₜ | 38.4 ± 0.47 | 0.937 ± 0.01 | 37.9 ± 0.57 | 0.933 ± 0.01 | 38.2 ± 0.48 | 0.936 ± 0.01 |

AuthorsOpenReview ↗

Response to Reviewer Gicz (Part 3/5)

Question 1: DDS baseline configuration.

Response: Thank you for pointing this out. We apologize for the lack of clarity in our description of the DDS configurations.

Given the denoised estimate x̂₀′ = E[x₀ | xₜ], DDS incorporates the measurement likelihood through a regularized least-squares data-consistency problem. The resulting linear system is solved using conjugate gradient (CG):

x̂₀ = CG(AᴴA + λI, Aᴴy + λx̂₀′, x̂₀′, k_CG).

Here, x̂₀′ is the estimate before CG correction, x̂₀ is the corrected estimate, and k_CG is the number of CG iterations.

DDIM and DPM++ can be combined with DDS by using the CG-corrected denoised estimate in their subsequent sampling updates. Below, t−1 denotes the next time point in the selected reverse-sampling schedule.

DDS-DDIM

xₜ₋₁ = sqrt(ᾱₜ₋₁) x̂₀ + sqrt(1 − ᾱₜ₋₁ − σₜ²) εθ(xₜ, t) + σₜz,

where z is sampled from N(0, I). The sampling noise scale is

σₜ = η sqrt[(1 − ᾱₜ₋₁)/(1 − ᾱₜ)] sqrt[1 − ᾱₜ/ᾱₜ₋₁].

For DDS-DDIM, we use η = 0.85 and retain the noise prediction computed before the CG correction.

DDS-DPM++

The first-order update using the CG-corrected denoised estimate is

xₜ₋₁ = sqrt(ᾱₜ₋₁) x̂₀ + sqrt[(1 − ᾱₜ₋₁)/(1 − ᾱₜ)] (xₜ − sqrt(ᾱₜ)x̂₀).

DDS Baseline

The configuration labeled “DDS Baseline” uses η = 1 and the full schedule of 1000 timesteps. This is a DDPM-style stochastic sampling variant with CG-based data-consistency corrections.

Regularization parameters

For MPI, we use λ = 100 at SNR = 20 dB and λ = 1 at SNR = 10 and 30 dB, except for DDS-DPM++ at SNR = 30 dB, where we use λ = 0.01. For MRI, we use λ = 10.

We have added detailed descriptions of the DDS configurations to Appendix A.4 to clarify these settings and avoid ambiguity.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses

We sincerely thank all reviewers for their careful reading and thoughtful, constructive comments. We appreciate their recognition of the potential of our approach, as well as their suggestions for improving its presentation and evaluation.

To address the reviewers’ concerns as thoroughly as possible, we devoted additional time to conducting supplementary experiments and analyses before finalizing our replies. We appreciate your patience and understanding.

The revised manuscript includes additional baseline comparisons and ablation studies, along with clearer explanations of the theoretical assumptions, training procedure, and sampling configurations. These revisions help distinguish our contributions from closely related work and clarify the scope of our claims.

The main updates are summarized below:

  • Additional baseline comparisons: We include recent score-based and flow-based inverse solvers, accelerated reconstruction methods, and a representative supervised unrolled baseline.
  • Detailed ablation studies: We examine the contributions of measurement conditioning, learned spatial weighting, and global equilibrium solving across different NFE budgets.
  • Clarified methodological foundations: We explain the assumptions underlying the posterior-score derivation, distinguish the learned weighting map from posterior covariance, and describe the training and data-consistency procedures more explicitly.
  • Further analysis of sampling efficiency: We compare the residuals and runtime of global and sequential solving under finite iteration budgets. We also evaluate reconstruction performance against advanced accelerated sampling methods across different NFE budgets.

We hope that these additions address the reviewers’ concerns and make the strengths and limitations of our method easier to assess. We remain grateful for the reviewers’ guidance and welcome further suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.