acceptodds
Under review as a conference paper at ICLR 2027

Decoupled Deep Unrolling Model for Snapshot Spectral Imaging

Abstract

Coded Aperture Snapshot Spectral Imaging (CASSI) compresses a 3D hyperspectral image (HSI) into a 2D measurement through cascaded mask modulation, spectral dispersion, and spectral integration. Existing deep unrolling methods typically represent these operations using a single coupled forward model and solve the resulting highly ill-posed reconstruction problem without dedicated regularization of intermediate physical states. To address these limitations, we propose the Decoupled Deep Unrolling Model (DDUM), which explicitly introduces intermediate state variables along the physical imaging chain of CASSI. Under a Bayesian formulation, DDUM decomposes reconstruction into three physically interpretable subproblems: integration inversion, spectral alignment, and mask demodulation. This decoupling enables dedicated regularization networks to model the distinct priors of different intermediate states, yielding more faithful reconstruction. To address the ambiguity inherent in many-to-one spectral integration, we further develop a learnable generalized physical inverse operator. We further analyze the inverse properties of the coupled and constituent operators. Extensive experiments demonstrate that DDUM consistently outperforms state-of-the-art methods and that adapting existing unrolling networks to our framework further improves their reconstruction performance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
hyperspectral image reconstruction, deep unrolling, inverse problems

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

This manuscript addresses hyperspectral image reconstruction in Coded Aperture Snapshot Spectral Imaging (CASSI). The authors propose DDUM, a physics-guided deep unrolling framework that decomposes reconstruction into three physically interpretable subproblems: integration inversion, spectral alignment, and mask demodulation. The framework incorporates intermediate-state priors and a learnable generalized physical inverse operator to improve reconstruction accuracy. Experiments on synthetic and real CASSI datasets are conducted to validate the proposed method.

Critical strengths
  1. The proposed decoupled framework is motivated by the physical imaging process of CASSI. By decomposing reconstruction into three inverse subproblems and introducing intermediate-state priors, the method provides a physically interpretable formulation for hyperspectral image reconstruction.

  2. The paper includes experiments on synthetic and real datasets. The reported improvements when applying the decoupling strategy to existing unrolling networks provide some evidence of its applicability to different architectures.

Critical weaknesses
  1. The novelty of the proposed architecture relative to MiJUN (AAAI 2025) is insufficiently justified. Both methods employ deep unrolling frameworks with U-shaped Mamba-based regularization networks and multi-scale feature extraction. In particular, the proposed SPAB in Fig. 4 appears structurally similar to the global-local attention mechanism in MiJUN, while the GeoT strategy shares the general objective of multi-directional feature modeling. Although DDUM introduces a distinct physically decoupled formulation, the manuscript does not clearly distinguish its network components from these existing designs or sufficiently establish their independent contributions.
  2. The experimental evidence is insufficient to fully support the claimed effectiveness of the decoupling strategy. Although MiJUN is included as a strong baseline in the quantitative comparisons, its reconstruction results are absent from the synthetic visual comparisons in Fig. 5, limiting direct qualitative assessment.While Table 5 shows performance improvements after introducing the decoupled formulation, the coupled and decoupled variants also differ in parameter counts and FLOPs. The stronger regularization network partially addresses the effect of model capacity, but the impact of additional computational overhead remains insufficiently examined.
Questions for authors
  1. The architectural novelty of DDUM relative to MiJUN (AAAI 2025) remains unclear. The proposed SPAB appears structurally similar to the global-local attention mechanism in MiJUN, while GeoT shares similarities with its multi-directional Mamba modeling.

  2. The ablation study in Table 7 only evaluates the removal of SPEB, SPAB, and MGF, without comparing them with similar modules from existing methods. Further comparisons with the corresponding designs in MiJUN are needed to better justify the independent contributions of these components.

  3. The coupled and decoupled variants in Table 5 differ in parameter counts and FLOPs. Although the authors include a stronger regularization network as a control, the influence of additional computational costs remains unclear. Further experiments under matched parameter and computational budgets are needed to isolate the contribution of the decoupling strategy.

  4. Although Mijun is included as a competitive baseline in the quantitative comparisons, its reconstruction results are absent from the synthetic visual comparisons in Fig 5. The authors should provide the corresponding visual results to facilitate a more comprehensive assessment of reconstruction details and artifacts.

5 comments
AuthorsOpenReview ↗

Response to Reviewer KtEk (part 1/2)

We sincerely thank the reviewer for the careful and constructive comments to improve the quality of our paper.

[Q1&Q2] Architectural Distinctions. DDUM's primary contribution lies in its physically decoupled reconstruction paradigm, rather than merely improving the Mamba architecture. As summarized in the Introduction, our contributions include: (1) Reconstruction Framework: decomposing the coupled CASSI problem into three physical inverse subproblems with dedicated intermediate-state priors; (2) Solution Mechanism: introducing a learnable generalized physical inverse to address the ambiguity of many-to-one spectral integration; and (3) Effectiveness and Generalizability: improving reconstruction performance across existing unrolling architectures. Notably, Table 5 shows that applying our decoupling framework to MiJUN's regularization backbone improves PSNR from 40.86 dB to 41.11 dB, demonstrating the effectiveness of our formulation without relying on DDUM's proposed regularizer.

Beyond the reconstruction formulation, DDUM also introduces distinct network designs:

1. SPAB vs. MiJUN's GLAM.

Structure: DDUM adopts an asymmetric U-Net, placing SPEB in the encoder and SPAB in the bottleneck and decoder, whereas MiJUN repeatedly employs GLAM-based MMIT blocks in both its encoder and decoder.

Role: Unlike MiJUN's repeated joint spatial–spectral modeling throughout its U-Net, DDUM explicitly separates spectral and spatial modeling: SPEB first exploits inter-band correlations for spectral recovery, while SPAB leverages global–local spatial dependencies to restore mask-degraded structures and fine details.

Results: Replacing SPAB with MiJUN's GLAM-based block reduces PSNR by 0.21 dB, as shown in the table below, demonstrating the effectiveness of our specialized spatial refinement design.

2. GeoT vs. MiJUN's Multi-Directional Mamba.

Structure: Unlike MiJUN's 12-direction mode-k scanning within each Mamba block, GeoT cyclically applies four geometric transformations across successive blocks, requiring only one SSM scan per block.

Advantage: By cyclically reorienting features across successive SSM blocks, GeoT captures complementary spatial–spectral dependencies without MiJUN's repeated multi-directional scans, achieving 0.10 dB higher PSNR in direct module comparisons comparison below.

3. Controlled Comparisons with MiJUN Modules. To further isolate the contributions of our network components, we replace individual DDUM modules with their MiJUN counterparts while keeping the decoupled reconstruction framework and other components unchanged.

ConfigurationPSNR (dB)Drop (dB)
SPAB → MiJUN's GLAM-based block41.220.21
GeoT → MiJUN's M-k Mamba41.330.10
SPEB → MiJUN's MMIT41.250.18
MGF → MiJUN's Fusion41.280.15
DDUM (Ours)41.43—

All four substitutions reduce reconstruction performance, directly demonstrating the effectiveness of DDUM's network components relative to their MiJUN counterparts, beyond the removal-based ablations in Table 7.

AuthorsOpenReview ↗

Response to Reviewer KtEk (part 2/2)

[Q3] Controlled Comparison. To isolate the benefits of our decoupled framework from increased model capacity or computational cost, we introduce two additional coupled baselines with identical parameters (0.68M) and FLOPs (72.17G):

1) Coupled + Three Priors: Retains three regularization networks but applies all of them to the final HSI.

2) Coupled + Enhanced Prior: Integrates the three networks into a single enhanced image-domain prior.

ConfigurationParams (M)FLOPs (G)PSNR (dB)
Coupled + Three Priors0.6872.1740.80
Coupled + Enhanced Prior0.6872.1740.83
DDUM (Ours)0.6872.1741.43

Under identical parameter and computational budgets, DDUM outperforms these coupled alternatives by 0.63 dB and 0.60 dB, respectively. These results demonstrate that the gains arise from decoupling-enabled physical-state regularization rather than increased model capacity or computational cost.

[Q4] Visual Comparison with MiJUN. MiJUN's original paper reports MiJUN-9stg (40.86 dB) as its best-performing configuration, which we therefore adopt for quantitative comparison. However, its published visualizations use MiJUN-5stg.

To provide a consistent comparison, following LRDUN (CVPR 2026), we have reproduced MiJUN-9stg using its publicly available code and added the visual results to Fig. 5 and Figs. 12–13 in the Appendix, while matching its reported quantitative performance. Some blurring and artifacts are observed, consistent with similar visual observations in LRDUN (Fig. 4). These additions enable direct qualitative comparisons under the same model configuration.

Reference: Huang, H., Guo, Y., & He, W. (2026). LRDUN: A Low-Rank Deep Unfolding Network for Efficient Spectral Compressive Imaging. CVPR 2026.

AuthorsOpenReview ↗

Gentle Reminder: Author Response Ready for Review

We sincerely appreciate your valuable time and constructive feedback! We hope our previous response and revisions have adequately addressed your concerns. As the discussion period is about to close very soon, we kindly invite you to review our response at your earliest convenience. We warmly welcome any remaining questions and would be delighted to provide further clarification before the deadline!

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

This paper proposes the Decoupled Deep Unrolling Model (DDUM) for coded aperture snapshot spectral imaging (CASSI) reconstruction. Instead of directly solving the inverse problem using a coupled sensing operator, DDUM introduces intermediate physical states and decomposes reconstruction into integration inversion, spectral alignment, and mask demodulation. The framework employs dedicated learnable priors for these intermediate states and introduces a generalized physical inverse operator to improve reconstruction. Experiments on synthetic and real CASSI measurements demonstrate promising results, including improved reconstruction quality and consistent gains when the decoupled formulation is applied to existing unrolling networks. However, several concerns regarding the theoretical analysis, methodological novelty, and robustness to physical model mismatch require further clarification.

Critical strengths
  1. Meaningful reconstruction formulation with physical interpretability.

The paper provides a structured approach to CASSI reconstruction by explicitly modeling intermediate states associated with the physical imaging process. Separating integration inversion, spectral alignment, and mask demodulation makes the reconstruction procedure more interpretable and enables the introduction of stage-specific regularization. This formulation offers an interesting alternative to conventional coupled deep unrolling frameworks.

  1. Promising reconstruction performance and cross-architecture applicability.

The experimental results are encouraging. DDUM achieves 41.43 dB PSNR on the synthetic benchmark, outperforming the evaluated baselines while maintaining a relatively compact parameter budget. More importantly, Table 5 shows that incorporating the decoupled formulation into DAUHST, DERNN, and MiJUN consistently improves reconstruction performance. These results provide useful evidence that the proposed framework is not limited to a particular backbone architecture.

Critical weaknesses
  1. The theoretical analysis of the coupled operator requires further justification.

The operator analysis in Table 1 and Appendix A is central to the motivation for decoupling, but some conclusions appear insufficiently justified. In particular, the appendix estimates the information retention ratio of the coupled operator as approximately p/Nλp/N_\lambda and claims that mask modulation strictly enlarges its null space whenever p<1p<1. However, Eqs. (38)–(40) show that AA⊤AA^\top is diagonal, with entries determined by the number of contributing unmasked spectral bands. Consequently, the rank of AA depends on the number of detector positions receiving at least one contribution, which does not necessarily decrease in proportion to pp. The claimed strict increase in null-space dimension therefore does not hold generally.

Furthermore, Appendix A assigns a computational complexity of O((Nλn)3)O((N_\lambda n)^3) to the coupled inverse problem based on naive dense matrix inversion. This does not account for the structured sensing operator. Since AA⊤AA^\top is diagonal under the authors' own formulation, the regularized inverse can be computed more efficiently using the corresponding measurement-space formulation. Therefore, the cubic complexity does not establish an inherent computational advantage of decoupling. The authors should revise these theoretical claims and clarify which benefits are supported by the analysis rather than by empirical results.

  1. Insufficient justification for the distinct benefits of intermediate-state priors over existing formulations.

The methodological novelty relative to existing deep unrolling approaches could be better established. DPU already incorporates separate image and degradation-residual priors, while LRDUN reformulates spectral compressive reconstruction using distinct low-rank variables and corresponding regularizers. DDUM instead introduces priors for the masked and dispersed intermediate states. However, the paper does not sufficiently explain what useful structures these particular representations expose, or why learning priors in these domains should be more effective than learning them directly in the image domain.

Table 5 demonstrates that the intermediate-state priors improve reconstruction performance, and the stronger image-prior baseline provides a useful capacity control. Nevertheless, these experiments do not fully distinguish the benefits of the physical-state representations from those of introducing additional regularization and iterative feature refinement. More targeted comparisons with parameter-matched multi-prior alternatives would strengthen the methodological contribution.

  1. Limited evaluation of physical model mismatch and spectral reconstruction fidelity.

The proposed intermediate-state priors depend on the assumed coded aperture and spectral dispersion operators. In practical CASSI systems, mask misregistration and dispersion calibration errors may alter the intermediate states on which these priors operate. It remains unclear whether the learned priors can tolerate such errors or may reinforce incorrect structures that propagate through subsequent reconstruction stages.

Although Figure 9 evaluates intermediate-state reconstruction errors and Table 6 examines robustness to additive Gaussian noise, neither directly addresses inaccuracies in the forward imaging model. Controlled experiments involving mask shifts and dispersion parameter perturbations would provide stronger evidence of practical robustness.

Additionally, Table 3 reports PSNR and SSIM but does not include a quantitative spectral distortion metric such as the mean Spectral Angle Mapper (SAM). Figure 8 presents spectral profiles for only two selected regions. Reporting SAM across the full synthetic test set would provide a more comprehensive assessment of spectral fidelity, which is particularly important for hyperspectral reconstruction.

Questions for authors

Q1. Could the authors clarify the derivation of the information retention ratio and null-space dimension of the coupled operator? Given that A⊤A^\top is diagonal under the stated model, why is the computational complexity of the coupled inverse characterized using cubic-cost dense inversion rather than a structured solution?

Q2. What specific structural properties make the masked and dispersed intermediate states more suitable for learning priors than the original image domain? Could the authors provide further controlled comparisons with parameter-matched multi-prior or image-domain regularization alternatives?

Q3. How robust is DDUM to mask misregistration and spectral dispersion calibration errors? Could the authors provide quantitative reconstruction results under such forward-model perturbations to establish whether the proposed intermediate-state priors remain effective?

4 comments
AuthorsOpenReview ↗

Response to Reviewer NSYE (part 1/3)

We thank the reviewer for the positive assessment and constructive comments.

[Q1] IRR and Nullity. Since the rank, nullity, and conditioning of the coupled CASSI operator depend on the specific mask configuration, spectral shifts, and boundary conditions, no universal exact characterization exists for all CASSI operators.

To characterize the coupled operator, we consider a structured CASSI model with a shared binary mask and integer circular shifts. Under these assumptions, Appendix A derives lower bounds on the IRR and null-space dimension, and an upper bound on the observable condition number, as summarized in Table 1. These bounds characterize the operator properties under the stated assumptions, given the unavailability of universal exact values.

Computational Complexity. As discussed in Appendix A.5, the original O((Nλn)3)O((N_\lambda n)^3) estimate assumes naive dense inversion of the coupled normal equations. Exploiting the diagonal structure of AA⊤AA^\top, the Woodbury identity reduces the data-consistency update complexity to O(N+ny)O(N+n_y), where N=NλnN=N_\lambda n.

However, despite Woodbury acceleration, fixed data-consistency updates may lead to inaccurate band-wise spectral corrections, degrading reconstruction accuracy (Liu et al., 2026). In contrast, DDUM leverages physical decoupling to introduce a learnable generalized physical inverse operator GUG_U, enabling adaptive, band-specific corrections to improve reconstruction fidelity.

We have revised Appendix A and Table 1 to clarify these derivations, theoretical bounds, and computational complexity.

Reference: Liu, S., Zhang, Y., Gu, L., Zhang, T., & Fu, Y. (2026). Manifold Prior Guided Deep Unfolding for Hyperspectral Image Reconstruction. NeurIPS 2026.

AuthorsOpenReview ↗

Response to Reviewer NSYE (part 2/3)

[Q2] Benefits of Physical-State Priors. Unlike DPU and LRDUN, which overlook the cascaded nature of CASSI's physical degradations, DDUM decomposes the coupled reconstruction into three operation-specific inverse subproblems, enabling dedicated regularization for each physical degradation.

Specifically, the masked state z=Mxz=Mx reveals mask-induced spatial structures, allowing R2(z)R_2(z) to regularize modulation-specific features. The dispersed state z′=Dzz'=Dz exhibits wavelength-dependent spatial shifts, enabling R3(z′)R_3(z') to regularize dispersion-specific structures.

Compared with learning priors directly in the image domain, stage-specific regularization improves intermediate-state fidelity, providing more reliable inputs for subsequent reconstruction stages and ultimately enhancing HSI reconstruction, as demonstrated in Table 5 and Fig. 9.

Controlled Comparison of Physical-State Priors. To further distinguish the benefits of physical-state regularization from simply adding more priors or increasing model capacity, we introduce two additional coupled baselines with matched parameters (0.68M) and FLOPs (72.17G):

1) Coupled + Three Priors: We retain all three regularization networks but apply them exclusively to the final HSI xx, controlling for the number of learned priors.

2) Coupled + Enhanced Prior: We integrate the three networks into a single enhanced image-domain prior Rnew(x)R_{\mathrm{new}}(x), controlling for the overall prior capacity under conventional coupled unfolding.

ConfigurationRegularizationParams (M)FLOPs (G)PSNR (dB)
Coupled + Three PriorsR1(x)+R2(x)+R3(x)R_1(x)+R_2(x)+R_3(x)0.6872.1740.80
Coupled + Enhanced PriorRnew(x)R_{\mathrm{new}}(x)0.6872.1740.83
DDUM (Ours)R1(x)+R2(z)+R3(z′)R_1(x)+R_2(z)+R_3(z')0.6872.1741.43

Under matched parameters and FLOPs, DDUM outperforms the two coupled baselines by 0.63 dB and 0.60 dB, respectively. These results demonstrate the advantage of physical-state regularization beyond simply increasing the number or capacity of image-domain priors.

AuthorsOpenReview ↗

Response to Reviewer NSYE (part 3/3)

[Q3] Physical-Model Mismatch. To evaluate the impact of calibration errors on DDUM, we extended Table 6 with mask misregistration (0.5-pixel spatial shift) and dispersion mismatch (a cumulative shift error of 0.5 pixel at the last spectral band), alongside different Gaussian noise levels.

Test ConditionDERNNLADEMiJUNOurs
Noise-free39.9340.0940.8641.43
Gaussian noise (σ = 0.025)39.5139.7240.5041.28
Gaussian noise (σ = 0.05)39.2339.2040.4441.21
Mask mismatch (0.5 pixel)32.5032.7632.6933.59
Dispersion mismatch (0.5 pixel)30.5830.3730.5030.97

Under mask and dispersion mismatches, DDUM achieves the highest PSNR, outperforming the strongest competing methods by 0.83 dB and 0.39 dB, respectively. These results demonstrate that DDUM maintains a reconstruction advantage under the tested physical-model perturbations, supporting the practical robustness of its intermediate-state regularization.

Spectral Reconstruction Fidelity. To further evaluate spectral accuracy, we compare the Spectral Angle Mapper (SAM) with representative state-of-the-art methods across all ten KAIST test scenes.

MethodPSNR (dB) ↑Mean SAM ↓
DAUHST38.367.44
RDLUF39.575.48
DERNN39.935.28
MiJUN40.862.92
DDUM (Ours)41.432.54

DDUM achieves the lowest SAM of 2.54, outperforming MiJUN (2.92) with a 13.0% reduction in spectral distortion. These results demonstrate that DDUM improves not only spatial reconstruction quality but also spectral fidelity.

AuthorsOpenReview ↗

Gentle Reminder: Author Response Ready for Review

We sincerely appreciate your valuable time and constructive feedback! We hope our previous response and revisions have adequately addressed your concerns. As the discussion period is about to close very soon, we kindly invite you to review our response at your earliest convenience. We warmly welcome any remaining questions and would be delighted to provide further clarification before the deadline!

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

This paper proposes a Decoupled Deep Unrolling Model (DDUM) for CASSI-based hyperspectral image reconstruction. Instead of directly solving the inverse problem associated with the complete sensing operator, the method decomposes the imaging process into mask modulation, spectral dispersion, and spectral integration. Two intermediate variables are introduced, and the resulting optimization problem is unfolded into a multi-stage neural network that jointly updates the reconstructed hyperspectral image and the intermediate states. The method achieves competitive reconstruction performance with relatively low computational complexity. Experiments also suggest that the proposed decoupling strategy can improve several existing reconstruction backbones.

Critical strengths
  1. A physically motivated and potentially novel unfolding architecture. By explicitly decoupling the CASSI imaging process into three physical operations and introducing intermediate states, the proposed framework provides a meaningful alternative to conventional deep unfolding methods. The architecture is more than a straightforward modification of existing reconstruction networks.

  2. Promising experimental performance and generality. The method achieves competitive reconstruction quality with a relatively compact model. More importantly, applying the decoupled framework to several existing reconstruction backbones consistently improves their performance, suggesting that the proposed strategy has broader applicability.

Critical weaknesses
  1. Questionable computational complexity analysis. The analysis in Appendix A.5 assumes that solving the original coupled inverse problem requires an expensive matrix inversion with cubic complexity. However, the CASSI sensing matrix has a special structure that enables efficient data-consistency updates using the Woodbury identity or measurement-domain inversion. Therefore, the claimed computational advantage of the proposed decoupling strategy may be overstated.

  2. Insufficient theoretical rigor. The information retention ratio, rank, and null-space analyses in Appendix A are based on assumptions that may not hold for general CASSI sensing operators. In particular, the rank and nullity depend on the actual sensing matrix and its boundary conditions. These theoretical claims need more rigorous justification.

  3. Limited validation of intermediate variables. A key feature of the proposed framework is the explicit estimation of two intermediate states, zkz^k and z′kz'^k, corresponding to different physical imaging operations. Although Fig. 9 provides an initial evaluation of their reconstruction errors, more comprehensive quantitative results and visual comparisons are needed to establish their physical fidelity and demonstrate the benefits of the proposed decoupling.

  4. Unclear relationship between the optimization formulation and network updates. Some updates are described as closed-form solutions, although they involve linearized approximations or learned operations. For example, the update of zz uses the regularization gradient evaluated at the previous iteration. The authors should clarify which steps exactly solve the corresponding subproblems and whether the resulting unfolding process has any theoretical convergence guarantees.

  5. Insufficiently controlled ablation studies. The proposed framework combines physical decoupling with multiple learned regularizers and network components. Although the reported ablations demonstrate performance improvements, they do not fully isolate the contribution of the decoupled formulation itself. Comparisons with conventional unfolding networks under matched parameter counts and computational budgets would better support the claimed architectural advantage.

  6. Limited evaluation of robustness and stability. The decoupled formulation relies on the assumed physical operators for mask modulation, dispersion, and spectral integration. Errors in these operators may propagate across the intermediate reconstruction stages. Further experiments under calibration errors, dispersion mismatch, and different measurement noise levels would help establish the stability and practical applicability of the proposed framework.

Questions for authors
  1. How do the theoretical complexity and ill-posedness analyses change when the special structure of the CASSI operator is considered?

  2. Can the authors provide stronger quantitative and visual evidence of the physical fidelity of zkz^k and z′kz'^k?

  3. Can the authors isolate the benefits of decoupling through comparisons with conventional unfolding under matched model capacity and computational cost?

  4. Which updates are exact optimization steps, and which are learned approximations? Are there any convergence guarantees?

  5. How robust is the proposed framework to measurement noise and calibration errors?

4 comments
AuthorsOpenReview ↗

Response to Reviewer SDMF (part 1/3)

We appreciate the reviewer's valuable and constructive comments.

[Q1] Computational Complexity. As stated in Appendix A.5, the original complexity estimate of O((Nλn)3)O((N_\lambda n)^3) assumes a naive dense solver for the coupled normal equations. Exploiting the CASSI-specific structure AA⊤=diag⁡(g)AA^\top=\operatorname{diag}(g), the Woodbury identity reduces the complexity to O(N+ny)O(N+n_y), where N=NλnN=N_\lambda n.

However, even with Woodbury acceleration, fixed data-consistency updates may lead to inaccurate band-wise spectral corrections, degrading reconstruction accuracy (Liu et al., 2026). In contrast, DDUM introduces a learnable generalized physical inverse operator GUG_U through physical decoupling, enabling adaptive, band-specific corrections to improve reconstruction fidelity.

Ill-posedness Analysis. Since the rank, nullity, and conditioning of the coupled CASSI operator depend on the specific mask configuration, spectral shifts, and boundary conditions, no universal exact characterization exists for all CASSI operators.

To rigorously assess the coupled operator's structural disadvantages, we compare its best-case detector coverage with the individual operators in the decoupled formulation. Specifically, under the shared binary mask and integer circular-shift model in Appendix A, even with full detector coverage (the best case for rank and nullity), the coupled operator retains an (Nλ−1)n(N_\lambda-1)n-dimensional null space. Table 1 further summarizes the derived bounds on its IRR, nullity, and observable condition number.

These results highlight the coupled operator's structural disadvantages compared with its decoupled components, motivating DDUM's physical decoupling and stage-specific regularization.

We have revised Appendix A and Table 1 to clarify the computational complexity, theoretical bounds, and their applicability.

Reference: Liu, S., Zhang, Y., Gu, L., Zhang, T., & Fu, Y. (2026). Manifold Prior Guided Deep Unfolding for Hyperspectral Image Reconstruction. NeurIPS 2026.

AuthorsOpenReview ↗

Response to Reviewer SDMF (part 2/3)

[Q2] Intermediate-State Fidelity. To further validate the physical fidelity of intermediate states, we added Appendix F with quantitative and visual comparisons between DDUM and the coupled baseline across all unrolling stages.

Visually, at Stage 3, the coupled baseline still exhibits severe blurring and artifacts in both intermediate states z and z', whereas DDUM already recovers clearer spatial structures and finer details.

Quantitatively, at Stage 8, DDUM improves PSNR/SSIM over the coupled baseline from 48.00 dB / 0.9920 to 49.45 dB / 0.9963 for z, and from 44.85 dB / 0.9745 to 48.45 dB / 0.9921 for the aligned z'.

These results, together with Fig. 9, further demonstrate the superior physical fidelity of DDUM's intermediate states estimates.


[Q3] Controlled Ablation Study. To isolate the benefits of decoupling, we introduce two additional coupled unfolding baselines under matched parameters (0.68M) and FLOPs (72.17G):

1) Coupled + Three Priors: We retain the three regularization networks but apply all of them to the final HSI xx, rather than to distinct physical states.

2) Coupled + Enhanced Prior: Following the conventional coupled formulation, we integrate the three regularization networks into a single enhanced image-domain prior Rnew(x)R_{\mathrm{new}}(x).

ConfigurationRegularizationParams (M)FLOPs (G)PSNR (dB)
Coupled + Three PriorsR1(x)+R2(x)+R3(x)R_1(x)+R_2(x)+R_3(x)0.6872.1740.80
Coupled + Enhanced PriorRnew(x)R_{\mathrm{new}}(x)0.6872.1740.83
DDUM (Ours)R1(x)+R2(z)+R3(z′)R_1(x)+R_2(z)+R_3(z')0.6872.1741.43

Under identical computational budgets, DDUM outperforms the two coupled baselines by 0.63 dB and 0.60 dB, respectively. These results demonstrate that the improvements stem from decoupling-enabled physical-state regularization rather than increased model capacity or computational cost.

Furthermore, Tab. 5 and Fig. 1(c) demonstrate the generalizability of our framework, yielding PSNR gains of 0.75, 0.58, and 0.25 dB when applied to DAUHST, DERNN, and MiJUN, respectively.

AuthorsOpenReview ↗

Response to Reviewer SDMF (part 3/3)

[Q4] Optimization Updates. By linearizing R2(z)R_2(z) at z(t)z^{(t)}, Eq. (15) gives the exact minimizer of the resulting quadratic surrogate corresponding to the original subproblem. Eq. (17) exactly solves the quadratic auxiliary-variable subproblem, while Eqs. (13) and (18) involve learned approximations. In Eq. (14), ∇R2(z(t))\nabla R_2(z^{(t)}) is a typo and should be ∇R2(z)\nabla R_2(z).

Convergence. Similar to many deep unrolling networks, DDUM employs learned regularizers for which theoretical convergence guarantees cannot be directly established (Fang et al., 2024). Nevertheless, Fig. 7 shows that PSNR consistently improves and gradually stabilizes as the number of unrolling stages increases, while Fig. 9 demonstrates decreasing reconstruction errors of the intermediate states. These results provide empirical evidence of stable and convergent iterative behavior.


[Q5] Robustness to Noise and Calibration Errors. We extended Table 6 to evaluate different Gaussian noise levels (σ = 0.025, 0.05), mask calibration errors (0.5-pixel spatial misalignment), and dispersion mismatch (cumulative shift error reaching 0.5 pixel at the last spectral band).

Test ConditionDERNNLADEMiJUNOurs
Noise-free39.9340.0940.8641.43
Gaussian noise (σ = 0.025)39.5139.7240.5041.28
Gaussian noise (σ = 0.05)39.2339.2040.4441.21
Mask mismatch (0.5 pixel)32.5032.7632.6933.59
Dispersion mismatch (0.5 pixel)30.5830.3730.5030.97

DDUM achieves the highest PSNR across all tested conditions, with only a 0.22 dB drop at σ = 0.05. Under mask and dispersion mismatches, DDUM outperforms the strongest competing methods by 0.83 dB and 0.39 dB, respectively. These results demonstrate DDUM's robustness to measurement noise and its consistent reconstruction advantage under calibration errors.

Reference: Fang, Z., Buchanan, S., & Sulam, J. (2024). What's in a Prior? Learned Proximal Networks for Inverse Problems. ICLR 2024.

AuthorsOpenReview ↗

Gentle Reminder: Author Response Ready for Review

We sincerely appreciate your valuable time and constructive feedback! We hope our previous response and revisions have adequately addressed your concerns. As the discussion period is about to close very soon, we kindly invite you to review our response at your earliest convenience. We warmly welcome any remaining questions and would be delighted to provide further clarification before the deadline!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Major Contributions, Strengths, and Responses to Concerns

Dear AC, SAC, and PCs:

We sincerely appreciate your time and effort in handling our submission.

To assist in decision-making, we provide a concise summary of our paper's strengths and the main improvements we have made in response to reviewer comments. Hope this is helpful and can reduce your review burdens.

1. Core Contributions

  • A Novel Physics-Decoupled Paradigm. DDUM reformulates CASSI reconstruction into three physically interpretable inverse subproblems, enabling dedicated regularization of intermediate physical states.
  • A Learnable Generalized Physical Inverse. Introduces adaptive, band-aware corrections to address the ambiguity of spectral integration and improve reconstruction fidelity.
  • Strong Performance and Generalizability. Achieves 41.43 dB PSNR with only 0.68M parameters, while improving three existing unrolling models by +0.25 to +0.75 dB.

2. Strengths Recognized by Reviewers

We are encouraged that the reviewers recognized the significance of our work:

  • Novelty and Physical Interpretability. Reviewer SDMF described DDUM as "a physically motivated and potentially novel unfolding architecture" and "more than a straightforward modification of existing reconstruction networks." Reviewers NSYE and KtEk also recognized its physically interpretable formulation.
  • Strong Performance and Broad Applicability. Reviewers SDMF and NSYE highlighted DDUM's promising reconstruction performance and cross-architecture applicability, while KtEk acknowledged the improvements achieved when integrating our framework into existing models.

3. Summary of Rebuttal and Improvements

We have carefully addressed the reviewers' concerns through theoretical clarifications, controlled experiments, and additional evaluations.

Reviewer(s)Main ConcernOur Response and Evidence
SDMF, NSYETheoretical rigor and complexityClarified operator assumptions, theoretical bounds, and efficient Woodbury-based updates; revised Table 1 and Appendix A.
SDMF, NSYE, KtEkIsolating the benefit of decouplingAdded two coupled baselines with identical parameters (0.68M) and FLOPs (72.17G). DDUM achieves +0.63/+0.60 dB higher PSNR.
SDMFIntermediate-state fidelityAdded stage-wise quantitative and visual comparisons. DDUM improves the PSNR of intermediate states z and z′ by +1.45/+3.60 dB, respectively.
SDMF, NSYERobustness and spectral fidelityAdded noise and calibration-mismatch evaluations. DDUM maintains the best reconstruction quality and reduces mean SAM by 13.0% compared with MiJUN.
KtEkNovelty relative to MiJUNAdded direct module comparisons (0.10–0.21 dB gains over corresponding MiJUN modules) and reproduced MiJUN-9stg visualizations.
SDMFOptimization and convergenceClarified exact versus learned updates, explicitly discussed convergence limitations, and provided empirical stability evidence.

4. Concluding Remarks

We believe these additional results substantially strengthen DDUM's central contribution. Most importantly, the matched-budget comparisons demonstrate that the improvements arise from physical decoupling rather than increased model capacity or computational cost. The intermediate-state fidelity analysis and consistent gains across existing architectures further validate the effectiveness and generality of our framework.

We sincerely thank the reviewers for their constructive feedback and appreciate your consideration of our submission.

Best regards,
The Authors of DDUM

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.