acceptodds
Under review as a conference paper at ICLR 2027

HPDiff: Zero-Shot Blind Hyperspectral Pansharpening with Degradation-Guided Diffusion Sampling

Abstract

Hyperspectral pansharpening reconstructs a high-resolution hyperspectral image (HSI) from a low-resolution HSI (LRHS) and a high-resolution panchromatic (PAN) image. Existing methods trained with simulated degradations often rely on predefined spatial point spread functions (PSFs) and spectral response functions (SRFs). When these assumptions mismatch unknown sensor degradations, the reconstruction may fail to satisfy both observations. To address this issue, we propose HPDiff, a zero-shot blind hyperspectral pansharpening method with degradation-guided diffusion sampling. HPDiff first estimates the unknown PSF and SRF from a single test pair by matching spatially degraded PAN with spectrally projected LRHS in a shared low-resolution PAN domain. The estimated operators then define LRHS/PAN data consistency gradients to guide the reverse sampling of a pretrained low-rank Diffusion Sampling. During joint reconstruction, HPDiff further refines the degradation parameters around their initial estimates, improving observation adaptation while retaining physical plausibility. Observation Prior Guidance (OPG) integrates LRHS spectral information, PAN spatial structures, and lightweight task-information priors to regulate spectral, spatial, and edge constraints during reverse sampling. A Coupled Spatial-Spectral Low-Rank Decoder (CSLD) jointly estimates interacting spatial and spectral factors to map diffusion details into a full-band HSI. Experiments on multiple benchmarks demonstrate improved PSNR, SAM, and ERGAS, together with better preservation of spatial structures and spectral consistency. The code will be made publicly available upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hyperspectral pansharpening, Zero-shot, Diffusion Model, Degradation

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_BPRfOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HPDiff, a zero-shot blind hyperspectral pansharpening framework leveraging pre-trained RGB diffusion priors without paired training data. Its pipeline combines adaptive PSF and SRF estimation via domain alignment, physical degradation, and Sobel edge-guided reverse diffusion sampling, and test-time self-supervised optimization using spatial-spectral low-rank factorization.

Critical strengths
  1. Cleverly leverages low-resolution domain alignment with Softmax center-clustering for PSF and soft-gated windowing with second-order differences for SRF, effectively preventing trivial or degenerate solutions in blind degradation estimation.

  2. Translates physical re-projection residuals and Sobel edge losses into posterior gradient guidance during sampling, successfully balancing rich texture generation with strict observation consistency and suppressing diffusion hallucinations.

  3. Bypasses direct high-dimensional pixel prediction by factorizing outputs into spatial and spectral matrix components with orthogonality constraints, reducing parameters while preserving spectral fidelity.

Critical weaknesses

1.Fails to compare against recent (2024–2026) zero-shot diffusion and generative hyperspectral pansharpening methods, such as: (1) Xiao, Jin-Liang, et al. "Hyperspectral pansharpening via diffusion models with iteratively zero-shot guidance." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025. (2) Xie, Pengwei, et al. "Dual Modality Prompted Diffusion Priors for Zero Shot Hyperspectral Pansharpening." arXiv preprint arXiv:2608.11748 (2026).

  1. This framework primarily concatenates standard techniques by applying a classical Sobel filter for edge guidance and conventional low-rank factorization for spectral decoding, making the technical novelty of the guidance pipeline somewhat incremental compared to prior zero-shot diffusion frameworks.

  2. Calculating backpropagation gradients at every diffusion sampling step in Stage 2 and fine-tuning parameters in Stage 3 inherently incur heavy runtime and memory overhead. Notably, the paper lacks a systematic discussion or empirical evaluation regarding computational efficiency.

  3. Under realistic striping or severe Gaussian noise, the parameterized PSF and SRF estimation easily gets trapped in local minima or overfits noise.

  4. Using a static PCA-derived matrix fails to model non-linear spectral mixing in complex scenes, while a fixed subspace rank lacks adaptability across varying scene complexities.

Questions for authors
  1. Can the authors provide quantitative results comparing HPDiff against recent baselines, along with exact runtimes in seconds or minutes and GPU memory usage for standard-resolution images?

  2. How do the estimated PSF and SRF parameters deviate in terms of root mean square error against ground truth under varying signal-to-noise ratio levels? Are there specific regularization mechanisms to prevent noise overfitting?

  3. Is the representative band extraction matrix updated dynamically during diffusion sampling? Furthermore, how sensitive are SAM and PSNR metrics to variations in the subspace rank across different land-cover complexities?

3 comments
AuthorsOpenReview ↗

Response to Reviewer BPRf

We thank the reviewer for the constructive comments. We address all five weaknesses and three questions below.

W1. Recent baselines. We would like to clarify that DM-ZS, already included in the submitted comparisons, is Xiao et al., "Hyperspectral Pansharpening via Diffusion Models With Iteratively Zero-Shot Guidance" (CVPR 2025). The correspondence is now explicit in Section 4.1. We also added the suggested DIDM method to the quantitative, qualitative, real data, and efficiency comparisons.

W2. Technical contribution. Sobel filtering and low rank factorization are established tools. Our contribution couples sample specific blind operator estimation, interacting observation conditions in OPG, and learned spatial-spectral decoding in CSLD. The estimated operators define reconstruction constraints; OPG couples PAN/LRHS information with task cues; CSLD exchanges spatial and spectral information through shared tokens. Table 3 supports their complementarity: combining OPG and CSLD improves PSNR over removing both by 1.826/1.383/0.865 dB on Pavia/Houston/Chikusei.

W3. Computational overhead. Section 4.6 and Table 7 now report full pipeline runtime, peak GPU memory, parameters, and FLOPs. We explicitly discuss the additional optimization cost and the performance-efficiency trade-off; numerical comparisons are given in Q1.

W4. Noise sensitivity. Appendix G adds operator RMSE measurements under Gaussian noise and an explicit regularization ablation (Q2). These quantify the dependence on noise level.

W5. PCA and spectral representation. HPDiff does not use PCA. Section 3.4 obtains the initial spectral lifting matrix by regularized least squares from selected LRHS bands; Section 3.6 learns the final spatial and spectral factors through CSLD. The decoder is nonlinear in its inputs, while its final synthesis remains rank constrained. We acknowledge the fixed rank as a hyperparameter and quantify its sensitivity in Q3.

Q1. Comparisons and computational costs. Tables 1-2 now include DIDM. Under known operators, HPDiff improves PSNR over DIDM by 0.909/0.330/0.863 dB on Pavia/Houston/Chikusei and ranks first in all 18 metric comparisons. Under shared estimated operators, it ranks first in 16/18 comparisons.

For a 256 × 256 Pavia image on an RTX 4090 with an i9-14900KF, Table 7 reports runtime/peak GPU memory of 313.934 s/7615.6 MB for HPDiff, 182.490 s/7201.9 MB for DIDM, 109.637 s/7601.2 MB for DM-ZS, and 39.076 s/6851.7 MB for PLRDiff. HPDiff's measurements include degradation estimation, OPG adaptation, all sampling, and CSLD optimization. Its runtime is 1.72 times DIDM's, with 5.7% higher peak memory. The additional cost supports degradation adaptation and reconstruction quality.

Q2. Operator errors and noise regularization. Table 11 reports results on three Pavia samples with two noise seeds, using unit-sum-normalized operators. For clean/30/20/10 dB conditions, the reported Default PSF RMSE means are 0.00016/0.00099/0.00200/0.00384, and SRF RMSE means are 0.00194/0.00291/0.00398/0.00479; standard deviations are also provided.

The estimator enforces non-negativity and normalization, with PSF smoothness and center priors and SRF second-order smoothness and window penalties. The No reg control removes the four explicit penalties while retaining the parameterization constraints. According to Table 11, regularization lowers operator RMSE in 7/8 comparisons. The effects at 10 dB are mixed: PSF RMSE and SAM improve, but SRF RMSE and PSNR worsen.

Q3. Matrix updates and rank sensitivity. Representative bands are selected at fixed intervals for each sample. Their indices and the initial lifting matrix E=YS⊤(SS⊤+ϵI)−1E=YS^\top(SS^\top+\epsilon I)^{-1} remain fixed within a sampling pass; no PCA is involved. CSLD parameters are subsequently optimized on the observed pair, and its learned factors are used for final reconstruction and refinement sampling. This differs from keeping the final spectral representation fixed throughout reconstruction.

The three diffusion channels are distinct from CSLD rank ρ\rho. Appendix C, Table 9 evaluates ρ=20/40/60/80/100\rho=20/40/60/80/100. The best tested ranks are 80/40/100 on Pavia/Houston/Chikusei for both PSNR and SAM. PSNR varies by 1.264/0.199/0.837 dB and SAM by 0.213/0.132/0.083° across the tested ranks. These results show scene-dependent sensitivity; we do not claim automatic rank selection or unrestricted modeling of physical nonlinear mixtures.

Reviewer_BPRfOpenReview ↗

Remaining Concerns on Novelty and Practical Value

Thank you for the response. The DM-ZS and spectral-representation clarifications resolve those points, but my concern about the technical contribution remains. The component ablations demonstrate benefits from their combination; they do not sufficiently establish what methodological advance distinguishes HPDiff from existing diffusion-guided reconstruction frameworks. Furthermore, the noise analysis covers only three Pavia samples with Gaussian noise and leaves the requested stripe-noise robustness untested. These additions improve the evaluation but do not sufficiently resolve my concerns. I maintain my original rating.

AuthorsOpenReview ↗

Response to Reviewer BPRf

Thank you for the follow-up. We clarify the methodological distinctions and expand the noise evaluation below.

1. Technical contribution.

We clarify the distinction of CSLD through its reconstruction mapping and the resulting diffusion guidance. Let AA denote the current denoised detail representation. With network parameters fixed during sampling, the separable NSSD used by DM-ZS/DIDM takes the form X=Fλ(Y)Fs(A)X=F_{\lambda}(Y)F_{s}(A). Although both networks are learned, the spectral factor depends only on LRHS, giving DAFλ=0D_{A}F_{\lambda}=0.

CSLD instead uses shared-token exchange to construct X=Fλ(A,Y,P)Fs(A,Y,P)X=F_{\lambda}(A,Y,P)F_{s}(A,Y,P). Both factors can therefore respond to changes in the current generated details, even while their network parameters remain fixed. The differential of the reconstruction is:

dX=Fλ,dFs+(dFλ)Fs.dX=F_{\lambda}\\,dF_{s}+(dF_{\lambda})F_{s}.

Consequently, with G=∇XLG=\nabla_{X}\mathcal{L}, the observation-guidance gradient becomes:

∇AL=(DAFs)∗[Fλ⊤G]+(DAFλ)∗[GFs⊤],\nabla_{A}\mathcal{L}=(D_{A}F_{s})^{\ast}[F_{\lambda}^{\top}G]+(D_{A}F_{\lambda})^{\ast}[GF_{s}^{\top}],

where the superscript star denotes the Jacobian adjoint. The second term is absent from the separable NSSD mapping. It allows observation residuals to influence the generated details through the spectral factor, complementing the spatial-factor feedback path. During refinement sampling, this gradient is propagated to the diffusion state through:

∇xtL=(DxtA)∗∇AL.\nabla_{x_{t}}\mathcal{L}=(D_{x_{t}}A)^{\ast}\nabla_{A}\mathcal{L}.

Thus, the distinction concerns how the reconstruction map feeds observation constraints back into diffusion, rather than low-rank factorization itself. The reported spectral-gradient comparison gives PSNR/SAM of 35.179 dB/5.937 degrees with full propagation versus 34.909 dB/6.036 degrees with spectral-factor feedback blocked; the other four aggregate metrics also favor full propagation.

The estimated PSF/SRF operators define the observation constraints used in this process, while OPG modulates the PAN/LRHS conditioning of the denoiser.

2. Noise robustness and regularization.

We expanded the evaluation from three to eight Pavia Center patches, using one noise seed and covering Gaussian and fixed-pattern stripe noise at 30/20/10 dB. Stripes affect 4% of columns per band with widths of 1-2 pixels; both noise types have equal total power at each SNR.

The table reports PSF/SRF RMSE against ground truth after unit-sum normalization. Default regularization combines PSF smoothness and center priors with SRF smoothness and window penalties. No reg removes the four explicit penalties while retaining the parameterization constraints. Regularization reduces operator RMSE in 13/14 comparisons, including all six stripe-noise comparisons. At 20 dB stripe noise, PSF/SRF RMSE decreases from 0.01390/0.02726 to 0.01074/0.01419.

Stage-1 reconstruction improves at 30 and 20 dB, although benefits become mixed at 10 dB. These results address the previously missing stripe-noise evaluation and support regularization for stabilizing estimation. They do not establish immunity to local minima or severe noise.

Reviewer_7BmSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a zero-shot, blind hyperspectral pansharpening framework called HPDiff. The core idea is to estimate sample-specific, effective PSF/SRF parameters from a given pair of PAN/LRHS observations and then incorporate these degradation operators into a diffusion-based reconstruction process. The authors map the result of applying spatial blurring and downsampling to the PAN image, as well as the result of applying spectral weighting to the LRHS image, into a common low-resolution PAN domain. They then optimize the PSF/SRF based on the consistency relationship D(B(P))≈R(Y)D(B(P)) \approx R(Y), where B(⋅)B(\cdot) denotes spatial blurring, D(⋅)D(\cdot) denotes spatial downsampling, and R(⋅)R(\cdot) denotes spectral weighting. This design ensures that, even in the absence of ground-truth HRHS data or known sensor parameters, the forward degradation model used for subsequent reconstruction is more accurately aligned with the specific observation pair at hand.

Critical strengths
  1. The problem is significant, and the motivation is clear. Instead of merely seeking a more powerful generative prior, the paper correctly identifies that diffusion-based posterior sampling relies on an accurate forward operator, and it explicitly addresses the issue of degradation mismatch, a perspective that is highly relevant to real-world sensor scenarios.

    1. The consistency constraints used for blind estimation are grounded in sound physical principles. Under the assumption that spatial degradation and linear spectral projection are commutative, Eq. (2) maps both the PAN and the LRHS into a common low-resolution PAN domain. Furthermore, the incorporation of specific priors—such as non-negativity, normalization, smoothness, and centrality for the PSF, as well as non-negativity, normalization, and continuous-support parameterization for the SRF, helps prevent ill-posed, unconstrained solutions for the degradation parameters.

    2. The proposed method establishes a comprehensive and well-structured reconstruction pipeline. There is a clear, logical connection between degradation estimation, diffusion guidance, the injection of observation conditions, and low-rank full-band decoding, rather than these being merely a collection of unrelated modules stacked together.

Critical weaknesses
  1. The paper claims the ability to "estimate unknown PSFs/SRFs." However, the current experiments primarily demonstrate that these estimated operators improve the final reconstruction; they do not directly verify the accuracy of the estimated PSFs/SRFs themselves. The authors propose a "blind" estimation of the PSF and SRF for a given sample using only a single PAN/LRHS pair, implying that the method not only yields superior reconstructed images but also successfully identifies the actual degradation. Yet, the data in Table 2 only show that using estimated PSFs/SRFs results in better reconstruction metrics compared to using random ones; it does not indicate how close the estimated PSFs/SRFs are to the ground truth. The authors are well-positioned to provide this verification: in their synthetic experiments, they generated data using known degradations—specifically, a 9 × 9 Gaussian kernel (σ = 1.6986) for the PSF and an averaging operation across all bands for the SRF. Since the ground-truth PSFs/SRFs are known, the authors could have directly compared the estimated PSFs/SRFs against the ground truth, but they failed to provide such a comparison.

  2. In Table 4, the "Proposed†" method (which utilizes estimated PSFs/SRFs within the HPDiff reconstruction architecture) is compared directly against rho-PNN*, PLRDiff*, and DM-ZS* (which use random degradation operators). Because the quality of the PSFs/SRFs varies across these different methods, the experimental conditions are not strictly equivalent. Table 2 indicates that replacing random PSF/SRF with estimated PSF/SRF leads to improvements in the reconstruction metrics for these baseline methods. Consequently, the superior performance of HPDiff shown in Table 4 may stem from two factors: the use of a more appropriate degradation operator, and the inherent effectiveness of its reconstruction approach. Since existing comparisons do not decouple these two aspects, it is not yet possible to definitively attribute the advantages of the HPDiff reconstruction framework to the framework itself, independent of the degradation operator. It is recommended that the authors evaluate all methods capable of incorporating PSF/SRF, specifically rho-PNN, PLRDiff, and DM-ZS, using the same set of estimated operators (denoted as rho-PNN†, PLRDiff†, and DM-ZS†). This would clarify whether the reconstruction component of HPDiff maintains its advantage when the degradation conditions are held constant.

3 In Section 3.5, the task cue q is described as "manually designed," yet the caption for Fig. 3 refers to it as "automatically generated", a clear inconsistency. If q is manually crafted for each test scenario, the method incorporates semantic information beyond the PAN/LRHS observation pairs, which complicates the interpretation of the zero-shot setting and raises concerns regarding the fairness of comparisons. If q is generated automatically, the authors should specify the generation process and the inputs used. Currently, the paper fails to provide the specific q values used for each dataset or sample, nor does it clarify whether the same cue is applied across all samples. Furthermore, while the paper mentions the use of a pre-trained RGB diffusion model as a spatial prior, it lacks details regarding the specific model architecture, the training data, and the checkpoint employed. Given that this diffusion model directly provides the spatial prior essential for reconstruction, the absence of these details hinders reproducibility and makes it difficult to discern whether performance gains arise from the proposed guidance mechanism or from the pre-trained model itself. The authors should explicitly state the origin of the task cue, the specific cues used in each experiment, and provide details on the pre-trained diffusion model (source, training data, and checkpoint). If the cues are indeed manually adjusted for different scenarios, the authors should provide additional comparisons—such as using a fixed, universal cue or omitting the cue entirely to demonstrate the impact of this supplementary information on the results.

Questions for authors
  1. In the synthetic setup, how close are the estimated PSF/SRF to the ground-truth values used to generate the LRHS/PAN data? Could you provide specific numerical values or a visual comparison?

    1. Why does Table 4 use estimated degradations (†) for HPDiff, while using random degradations (*) for rho-PNN, PLRDiff, and DM-ZS? Would the conclusions still hold if all these methods used the same estimated operators?

    2. What is the exact source of the "task cue" in Fig. 3? Is it provided manually or generated automatically? Are different cues used for different datasets or images? Is any semantic information beyond the observed pairs utilized?

    3. Could you specify the pre-trained diffusion model used, including the training data and the specific checkpoint? Is there any overlap between the pre-training data and the test scenarios or datasets?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 7BmS

We thank the reviewer for highlighting the need to distinguish operator fidelity, reconstruction quality, and the sources of prior information.

W1. Direct verification of PSF/SRF estimation. We agree that reconstruction gains alone do not verify parameter accuracy. Section 3.3 now explicitly describes the estimates as observation consistent effective operators, without claiming unique recovery of the physical degradation. Appendix F adds direct comparisons with the generating PSFs/SRFs, including kernel maps, response curves, and numerical errors (Q1).

W2. Fairness of the real data comparison. Revised Tables 5-6 evaluate rho-PNN, PLRDiff, DM-ZS, DIDM, and HPDiff under shared preset operators and shared HPDiff-estimated operators, respectively. These comparisons assess reconstruction under matched degradation conditions and separately show the benefit of operator adaptation (Q2).

W3. Task cue provenance and reproducibility. We corrected the inconsistent automatic/manual descriptions: cues are drafted by an LLM and manually checked. Appendix B explains their preparation and use; Appendix A identifies the pretrained prior. Table 4 adds a fixed Generic cue control. Relative to Generic, full OPG improves PSNR by 0.129/0.011/0.083 dB on Pavia/Houston/Chikusei, respectively, while reducing RMSE and SAM. These modest gains quantify the contribution of scene dependent cues.

Q1. How close are the estimated operators to ground truth? Appendix F (Figs. 8-9) compares estimates over eight Pavia samples with nine hidden PSFs and three SRF shapes. PSF cosine similarities are 0.789–0.990; for the nominal 9 × 9 Gaussian, cosine similarity is 0.931 and estimated/true effective widths are 1.30/1.65. Blur structures and orientations are largely preserved, although widths are underestimated by approximately 11%–22%.

The additional SRF tests use a shared window (bands 5-92, zero-based). For uniform, Gaussian, and skewed responses, RMSE values are 0.0074/0.0065/0.0063 and support IoUs are 0.896/0.864/0.811. Spectral integration allows compensating band weights to produce similar PAN projections, so local response oscillations can remain weakly constrained. Regularization stabilizes these effective responses. The comparisons characterize physical agreement, while reconstruction results assess practical utility.

Q2. Does the advantage persist with shared estimated operators? Yes. In revised Table 6, all five learning-based methods use HPDiff-estimated operators. HQNR is 0.699/0.729/0.704/0.682 for rho-PNN/PLRDiff/DM-ZS/DIDM, versus 0.738 for HPDiff; HPDiff also has the lowest spatial distortion (DS=0.196D_S=0.196). Under shared preset operators (Table 5), HPDiff still leads with HQNR 0.725 versus PLRDiff's 0.716. Estimated operators improve HQNR for every method. These results support contributions from both reconstruction and degradation adaptation.

Q3. What is the source and role of the task cue? PAN/LRHS visualizations are supplied to GPT 5.6 Terra to draft scene and land cover descriptions, followed by manual checks using the observed pair. No HRHS ground truth or external land cover materials are consulted. Cues are held fixed during reconstruction; they are not one universal prompt across datasets. We explicitly acknowledge human assisted semantic conditioning; zero shot refers to the absence of paired target scene training data.

Appendix B provides the complete cue and parsed weights for the illustrated Pavia sample. A keyword parser maps text to category weights, conditioning tokens, spectral band weights, and guidance coefficients. The Generic control uses a fixed scene independent description to assess the effect of scene specific semantics (Table 4).

Q4. Pretrained model, checkpoint, and test data separation. We use Bandara et al.'s DDPM-CD prior (WACV 2025), a frozen RGB pixel-space U-Net with three input/output channels. Checkpoint I190000_E97_gen.pth matches the official diffusion-model-I190000_E97_gen.pth. The diffusion comparisons share this pretrained backbone.

The source paper reports approximately 500,000 Sentinel-2 RGB patches from 200,000 locations; Appendix A also lists the authors' other reported training sources.

The prior was pretrained on three-channel RGB imagery, whereas our test targets are patches cropped from original multiband hyperspectral scenes. Geographic overlap alone does not establish reuse of identical observations: RGB and hyperspectral data differ in spectral representation, and their spatial resolution and acquisition conditions may also differ. We do not use HRHS targets to train or fine-tune the prior; adaptation uses only PAN/LRHS observations without HRHS supervision. Our protocol therefore prevents leakage through HRHS target supervision.

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

This paper proposes HPDiff, a zero-shot blind hyperspectral pansharpening method that integrates degradation estimation with diffusion-based reconstruction. Specifically, the method estimates unknown spatial and spectral degradation operators from the input observations and incorporates them into diffusion sampling to improve reconstruction consistency. It further introduces observation prior guidance and a coupled spatial-spectral low-rank decoder to enhance spatial details and spectral fidelity. Experimental results on multiple datasets demonstrate promising reconstruction performance. However, several concerns regarding the realism of degradation settings, fairness of comparisons, computational efficiency, and real-world generalization require further clarification.

Critical strengths
  1. Meaningful problem formulation with practical relevance.

The paper addresses the important problem of hyperspectral pansharpening under unknown sensor degradations. Estimating sample-specific PSF and SRF operators without requiring paired training data or accurate sensor calibration is practically valuable, particularly when degradation characteristics vary across sensors and acquisition conditions. The proposed observation-consistency formulation provides a physically motivated approach to this problem.

  1. Effective integration of complementary reconstruction components.

The proposed framework combines degradation estimation, observation-guided diffusion sampling, and spatial-spectral low-rank reconstruction in a coherent pipeline. The experimental results show consistent improvements in reconstruction quality across multiple datasets, while the ablation studies demonstrate the complementary contributions of observation prior guidance and the low-rank decoder. These results provide encouraging evidence for the effectiveness of the overall framework.

Critical weaknesses
  1. Limited realism of degradation settings and insufficient real-world validation.

Section 4.1 employs a fixed Gaussian blur kernel and uniform spectral band averaging to generate synthetic degradations, which may be overly idealized for a method targeting unknown sensor degradations. Experiments with non-Gaussian PSFs, nonuniform SRFs, and spatial misregistration would provide stronger evidence of generalization. Furthermore, the real-world evaluation is limited to a single 240×240240\times240 PRISMA image, making it difficult to assess robustness across different sensors and acquisition conditions. Although HPDiff achieves the highest HQNR in Table 4, its spectral distortion (Dλ=0.0940D_\lambda=0.0940) is higher than that of several baselines. Additional real-world experiments and a more detailed analysis of the trade-off between spatial reconstruction and spectral fidelity are needed to support the claimed practical effectiveness.

  1. Concerns about comparison fairness and computational efficiency.

In Table 4, several baselines use randomly specified degradation operators, whereas HPDiff uses estimated ones. Although Table 2 partially addresses this concern by applying the estimated operators to competing methods, the real-world comparison does not clearly separate the benefits of degradation estimation from those of the reconstruction framework. More controlled evaluations under identical degradation operators would strengthen the comparison. Additionally, HPDiff involves extensive test-time optimization, including degradation estimation, condition adaptation, and low-rank decoder optimization, but its computational efficiency is not evaluated. The authors should report runtime, peak GPU memory consumption, and efficiency comparisons with existing methods. Without these results, it is difficult to assess whether the reconstruction improvements justify the additional computational cost.

Questions for authors

Q1. Since the synthetic experiments use a fixed Gaussian PSF and uniform spectral averaging, how well does HPDiff generalize to non-Gaussian PSFs, nonuniform SRFs, and spatial misregistration? Can the authors provide additional results under more realistic degradation conditions?

Q2. In Table 4, some baselines use randomly specified degradation operators, whereas HPDiff uses estimated ones. Although Table 2 partially addresses this issue, can the authors clarify whether the real-world performance gains mainly arise from degradation estimation or the proposed reconstruction framework? Additionally, how do the authors explain the higher spectral distortion (DλD_\lambda) of HPDiff compared with several baselines?

Q3. Given the extensive test-time optimization involved in HPDiff, what are its runtime and peak GPU memory consumption per image? Can the authors provide computational efficiency comparisons with existing methods to demonstrate whether the reconstruction improvements justify the additional computational cost?

1 comment
AuthorsOpenReview ↗

Response to Reviewer rDDb

We thank the reviewer for the constructive comments. The revision adds controlled comparisons and explicitly discusses the remaining limitations.

W1. Realism and real world validation. We added non-Gaussian PSFs, nonuniform SRFs, PAN misregistration, and combined degradations on eight Pavia Center samples (Appendix E, Table 10; details in Q1). Real data evaluation now covers 10 non-overlapping patches from the PRISMA FR1 scene, with PAN/LRHS sizes of 240 × 240/40 × 40, without synthetic degradation. This expands spatial coverage within FR1; it does not establish generalization across sensors or acquisition conditions. Section 4.5 now explicitly discusses the spatial-spectral trade-off: HPDiff achieves the highest HQNR and lowest spatial distortion, but not the lowest spectral distortion (Q2).

W2. Comparison fairness and computational efficiency. We apologize for an editing error in the original Table 4: the baselines used preset operators, whereas HPDiff used estimated operators; describing the baseline operators as random was incorrect. This comparison mixed the effects of operator estimation and reconstruction. We therefore redesigned the evaluation: rho-PNN, PLRDiff, DM-ZS, DIDM, and HPDiff share preset operators in Table 5 and HPDiff-estimated operators in Table 6 (Q2). Section 4.6 and Table 7 now report runtime, peak GPU memory, parameters, and FLOPs, and discuss the accuracy-efficiency trade-off (Q3).

Q1. Generalization to realistic degradations. The eight added conditions use identical blind reconstruction settings and CSLD rank 20. Anisotropic Gaussian, motion, and disk PSFs yield PSNR of 34.20-34.84 dB and SAM of 6.04-6.36°. Gaussian and skewed SRFs yield 34.74/35.03 dB and 5.88/5.85°, respectively. These results support applicability beyond the original Gaussian/uniform setting for aligned observations. The main table reference rows retain different configurations, so their differences from these results are not attributed solely to degradation changes.

Misregistration remains a limitation. PAN shifts of (2,2)/(3,3) pixels yield 22.57/21.42 dB PSNR and 9.78/11.26° SAM; the combined anisotropic PSF, skewed SRF, and (2,2) shift yields 22.63 dB. Shifted PAN structures conflict with LRHS constraints because corresponding locations no longer represent the same content, and HPDiff has no explicit offset compensation. Metrics include zero filled borders. Registration across sensors is already identified as future work in the Conclusion.

Q2. Sources of the gains and spectral distortion. Both reconstruction and degradation adaptation contribute. With shared preset operators, HPDiff achieves HQNR 0.725 versus 0.716 for PLRDiff, the strongest learning-based baseline. Sharing estimated operators improves HQNR for all five methods; HPDiff and PLRDiff reach 0.738 and 0.729, respectively. The persistent advantage under matched operators supports the reconstruction framework, while the improvement between protocols supports degradation adaptation.

A plausible explanation for the higher DλD_\lambda lies in HPDiff's PAN fidelity and edge guidance. PAN constrains only a weighted spectral sum, while LRHS constrains blurred and downsampled spectra; neither uniquely specifies the spectrum at each high-resolution pixel. PAN-guided detail enhancement can therefore alter individual band intensities in ways that the soft LRHS constraint does not fully suppress. HPDiff has lower DSD_S than PLRDiff under both protocols (0.201/0.196 versus 0.235/0.235), but higher DλD_\lambda (0.098/0.081 versus 0.064/0.046). Estimated operators reduce both distortions for HPDiff, but do not eliminate the spectral gap. Its higher HQNR reflects the combined spatial and spectral outcome; it does not imply uniformly superior spectral fidelity.

Q3. Runtime, memory, and the performance-efficiency trade-off. Table 7 reports per-image costs on Pavia Center at 256 × 256 pixels, using an RTX 4090 and Intel i9-14900KF. HPDiff takes 313.934 s (5.23 min) with 7615.6 MB peak GPU memory, covering degradation estimation, OPG adaptation, all sampling, and CSLD optimization. Runtime/peak memory are 9.265 s/113.8 MB for rho-PNN, 39.076 s/6851.7 MB for PLRDiff, 109.637 s/7601.2 MB for DM-ZS, and 182.490 s/7201.9 MB for DIDM.

HPDiff requires 1.72 times the runtime of DIDM, with 5.7% higher peak memory. Its memory usage is comparable to DM-ZS. For context, under shared estimated operators HPDiff improves Pavia PSNR over DIDM by 1.088 dB (Table 2) and FR1 HQNR by 0.056 (Table 6). These results quantify the additional cost alongside the reconstruction benefit. HPDiff prioritizes degradation adaptation and reconstruction quality at higher computational cost.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.