acceptodds
Under review as a conference paper at ICLR 2027

Self-Supervised Variational Diffusion with Temporal Tucker Modeling for Hyperspectral Image Restoration

Abstract

Hyperspectral image (HSI) restoration aims to recover spatial structures and spectrally faithful signatures from measurements affected by noise, missing observations, and spatial degradation. Existing supervised methods often generalize poorly across sensors and corruption distributions, while model-based approaches rely heavily on manually designed priors with limited expressive ability. Diffusion-based HSI restoration provides a promising alternative, but existing methods remain limited in modeling complex hyperspectral corruption and in fully exploiting the regularization imposed by diffusion sampling and HSI-specific priors. To address these limitations, we propose VITTA, a self-supervised ariational dffusion framework with emporal ucker modeling for HSI restortion. Specifically, VITTA couples a likelihood step and a prior step within the reverse diffusion process. In the likelihood step, a two-factor probabilistic model characterizes complex HSI corruption, and the inferred non-Gaussian uncertainty is projected into Gaussian observation channels to provide reliability-aware measurement conditioning. Complementarily, the prior step employs a Temporal Tucker Prior Module to estimate the clean endpoint under joint spatial, spectral, and reverse-step regularization. The resulting likelihood and prior information are then integrated through reliability-guided conditional diffusion sampling to progressively restore the HSI. Theoretically, we establish approximation guarantees and recursive error bounds, verifying the feasibility of the proposed progressive recovery process. Extensive experiments demonstrate that our proposed method outperforms state-of-the-art methods on various tasks including HSI denoising, band completion and pansharpening.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Hyperspectral Image Restoration, Remote Sensing, Diffusion Model

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_kUGaOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies hyperspectral image (HSI) restoration under complex corruptions in a self-supervised setting. Building on the use of diffusion models for inverse problems, the authors aim to improve the ability of self-supervised diffusion methods to handle complex corruptions in HSI. To this end, the authors propose VITTA, which combines variational inference of observation noise and its uncertainty with a Temporal Tucker Prior Module to improve restoration performance under complex corruptions. The authors then theoretically analyze error bounds of the proposed VITTA method and demonstrate better restoration performance than the compared methods in experiments involving complex corruptions.

Critical strengths
  1. Although extending self-supervised diffusion methods to complex corruption settings is relatively incremental, these settings represent practical problems commonly studied in HSI restoration. Moreover, the authors consider the structural differences between HSI data and ordinary images and introduce a Temporal Tucker Prior Module to provide prior information about clean HSI data. This makes the proposed method more than a straightforward application of existing diffusion methods.
  2. As a diffusion-based method for inverse problems, the authors provide a theoretical analysis of the proposed restoration method. Under certain assumptions, they establish an approximation result and a conditional error bound. In the context of inverse problems, these analyses help explain the performance of the proposed method.
Critical weaknesses

Critical Weaknesses

  1. In diffusion-based restoration, existing studies have proposed various mechanisms for modeling complex noise and sparse outliers in observations, including approaches based on non-Gaussian likelihoods. This paper instead adopts a Gaussian projection, motivated by the need to "adapt the non-Gaussian observation residual for Gaussian-formulated conditional diffusion." If this design is intended as a contribution, the authors should more clearly explain and validate its relative advantages over directly using robust non-Gaussian observation models.
  2. Limited experimental coverage. The paper demonstrates the performance of VITTA under fixed complex degradation settings, but does not verify whether its advantages hold under stronger or milder noise conditions, as the tested signal-to-noise ratio (SNR) range is relatively narrow. Moreover, the mixed-corruption experiments do not cover certain cases, such as Gaussian-only noise or high proportions of impulse noise. It would be important to examine the method's performance under these conditions and their effects on parameter selection, as these settings are directly related to the proposed noise modeling mechanism.

Minor concerns

  1. The method involves several hyperparameters, including different numbers of diffusion steps K for different tasks. Given the variation and precise settings of these hyperparameters, the authors should clarify the criteria used for hyperparameter selection and the associated tuning cost.
  2. In some settings in Table 1, adaptive methods outperform pretrained supervised models. The authors should further clarify whether the training and testing settings of different methods are comparable, particularly regarding how well the pretrained models match the noise distributions used in testing and how their checkpoints were selected.
Questions for authors

My questions are those raised in critical weaknesses 1 and 2 above, they are the main concerns affecting my assessment.

2 comments

Thanks.

Response to CW 1

Compared with directly using robust non-Gaussian observation models, we would like to clarify that the advantage of VITTA lies not in approximating non-Gaussian noise as Gaussian, but in integrating explicit non-Gaussian corruption inference with reliability-guided Gaussian conditional diffusion.

VITTA infers band-dependent dense noise and sparse heavy-tailed corruption via variational inference, then applies forward-KL-optimal Gaussian projection to turn uncertainty into adaptive observation reliability. This preserves first-order and second-order statistics and quadratic risk of the affine transition, while bounding higher-order discrepancies that diffusion further attenuates in reliable channels. These advantages are further reflected in comparisons with existing restoration methods. For example, neither HLRTF nor GNRSR explicitly incorporates probabilistic noise uncertainty into diffusion sampling, whereas VITTA exploits the stronger generative priors and progressive refinement capabilities of diffusion sampling, further enhanced by temporal regularization across reverse steps. Additionally, we replaced VITTA's variational noise-adaptation component with Huber and IRLS [1], obtaining average PSNR/SSIM/SAM results of 31.254/0.8375/0.1025 (Huber) and 30.655/0.8118/0.1073 (IRLS), with 33.138/0.8911/0.0957 for VITTA. These approaches lack explicit modeling of different noise components and their associated uncertainty, resulting in being inferior to VITTA. These results demonstrate VITTA's superiority under complex degradations.

[1] Outlier-Robust Diffusion Solvers for Inverse Problems

Response to CW 2

We conducted additional experiments as required to verify VITTA’s robustness.

  • Weaker and stronger mixed noise: We scaled the complete corruption residual according to yα=x+α(yoriginal−x)y_{\alpha}=x+\alpha(y_{\mathrm{original}}-x), where xx denotes the clean HSI (downsampled for pansharpening), yoriginaly_{\mathrm{original}} is the observation generated under the original mixed-corruption protocol, and α\alpha controls the overall corruption intensity. Specifically, we set α=0.5\alpha=0.5 and α=1.5\alpha=1.5.

  • Gaussian-only noise: Independent Gaussian noise with σ=0.1\sigma=0.1.

  • High-density impulse noise:5050\\%–7070\\% corruption ratios.

Results in Table R1 confirm VITTA’s robustness. Due to space limitations, more details are in Appendix L.5.

Table R1.

Best Baseline→VITTA: PSNR/SSIM/SAM.

NoiseDenoise/ChikuseiCompletion/Houston18Pansharpening/PaviaU
Mix α=0.536.215/0.9257/0.0479→36.352/0.9510/0.045435.578/0.9010/0.1166→35.973/0.9126/0.100028.993/0.8350/0.1666→29.660/0.8381/0.1342
Mix α=1.529.855/0.7947/0.0957→31.870/0.8660/0.075029.344/0.6701/0.2590→31.788/0.8041/0.167925.955/0.7624/0.2753→26.374/0.7725/0.2136
Gauss σ=0.136.546/0.9355/0.0468→37.215/0.9564/0.041634.447/0.8746/0.1222→35.690/0.9075/0.099829.625/0.8393/0.1481→29.682/0.8403/0.1428
Impulse 50–70%31.066/0.8105/0.0771→31.876/0.8758/0.093131.423/0.7678/0.1952→32.509/0.8318/0.166526.575/0.7781/0.2500→27.506/0.7355/0.2012

Different degradation levels also affect VITTA's parameter selection since varying degradations require different model capacities and regularization strengths to balance noise suppression with detail preservation. For instance, increasing mixed-noise intensity shifts λtraj\lambda_{\mathrm{traj}} from 0.01 to 0, while raising impulse corruption adjusts Tucker ranks (Rs,Rλ,Rtemp)(R_s, R_{\lambda}, R_{\mathrm{temp}}) from (40,16,4)(40, 16, 4) to (40,16,2)(40, 16, 2). Notably, VITTA selects these parameters via no-reference NIQE and Spectral Roughness without ground truth.

Response to CW 3

VITTA selects its parameters using the no-reference metrics NIQE and Spectral Roughness without GT. Since different scenes and degradation types require different model capacities and regularization strengths, our adaptive selection strategy is reasonable and consistent with practices widely adopted in prior studies. The associated tuning cost requires approximately 16 trials on average to determine the final parameter settings. We also tuned competing methods to achieve their best possible performance.

Response to CW 4

All methods are evaluated under identical testing conditions for fairness. Importantly, the noise distributions used to train pretrained models are broadly consistent with our testing degradations, covering Gaussian noise and sparse corruptions such as impulse noise, stripes, and dead lines. For example, SERT explicitly includes non-i.i.d. Gaussian noise, Gaussian noise combined with stripes, dead lines, or impulse noise, and mixed-noise cases during training. We use the official pretrained checkpoints. Although specific noise intensities and proportions may differ, such variations are common in real HSI applications, whereas VITTA can adapt directly to the corruption characteristics of each observed scene.

Thank you for the detailed responses and additional experiments. The responses have adequately addressed my concerns.

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

This paper proposes VITTA, a self-supervised variational diffusion framework for hyperspectral image restoration under complex degradations. The method extends measurement-conditioned diffusion restoration by incorporating a two-factor variational noise model, Gaussian projection for reliability-aware observation conditioning, and a Temporal Tucker Prior Module for spatial–spectral representation and reverse-step regularization. The authors further provide theoretical analyses of the projection and sampling procedures and report improvements over various baselines on HSI denoising, missing-band completion, and noisy pansharpening.

Critical strengths

The reported empirical performance is promising. VITTA achieves consistently strong quantitative results across three HSI restoration tasks and multiple datasets, with improvements over several established optimization-based and learning-based baselines. The evaluation also includes visual comparisons on real HSI observations, which helps demonstrate the potential practical relevance of the method.

The paper provides a relatively complete methodological description and extensive supplementary material. The authors document the variational inference procedure, conditional sampling mechanism, temporal representation, theoretical assumptions, optimization settings, and computational costs. The component and loss ablations also provide some initial evidence that the proposed modifications contribute to the final performance. Overall, I acknowledge the effort involved in developing and evaluating this framework. Nevertheless, I remain unconvinced that the current contributions are sufficiently distinctive or well justified for acceptance. In particular, the relationship between the added methodological complexity and the actual restoration benefits deserves a more careful investigation.

Critical weaknesses

The methodological novelty and task-specific motivation are not sufficiently convincing. The overall framework largely follows existing measurement-conditioned diffusion restoration paradigms, particularly DDS2M and DDRM, while the main additions (variational noise modeling, Gaussian projection, and temporal Tucker regularization) are individually established techniques. The paper does not clearly explain why this particular combination is necessary or specifically tailored to hyperspectral remote sensing. For instance, the measurement-wise corruption model does not explicitly capture the spatial dependencies of stripe and dead-line artifacts, while the cosine-based temporal parameterization lacks a compelling justification over simpler alternatives. Moreover, despite the extensive theoretical analysis, the connection between the derived guarantees and the actual adaptive restoration process remains unclear. The moment-preservation results apply to restricted settings, and the recursive convergence analysis relies on assumptions and residual terms that are not empirically verified. Overall, the method appears more like an integration of existing techniques than a framework motivated by a sufficiently distinctive technical insight.

The experimental evidence is insufficient to establish a convincing advantage over existing methods, particularly considering the substantial computational overhead. Several relevant baselines are missing, including Diff-Unmix (CVPR 2024). More importantly, MP-HSIR is evaluated using Natural_scene.ckpt, despite its official implementation providing a dedicated remote-sensing checkpoint, raising concerns about comparison fairness. The evaluation also relies primarily on a small number of fixed image crops and a custom synthetic degradation protocol, without sufficient statistical analysis across independent corruption realizations or clear justification of dataset-specific hyperparameter choices. Furthermore, the ablations mainly examine sequential component additions and do not adequately isolate the contributions of the variational likelihood, temporal regularization, and diffusion sampling against simpler alternatives. Finally, VITTA requires hundreds to thousands of reverse steps, with reported restoration times frequently exceeding 1000 seconds per image. Without controlled comparisons against simpler optimization strategies or a meaningful quality–runtime trade-off analysis, it remains unclear whether the performance improvements justify the additional complexity and computational cost.

Questions for authors

Compared with DDS2M, what is the key methodological insight that fundamentally motivates the proposed framework, beyond incorporating a more elaborate likelihood model and structural regularization? Could the authors identify which design choices are specifically motivated by the characteristics of hyperspectral remote sensing?

The corruption model in Eqs. (6)–(12) treats sparse corruption through measurement-wise latent states and heavy-tailed amplitudes. How does this formulation explicitly capture spatially correlated artifacts such as stripes and dead lines? Moreover, the variational inference estimates a potentially nonzero sparse-corruption mean, whereas the Gaussian projection uses a zero-centered auxiliary residual distribution. Could the authors clarify the rationale for this choice and its effect on restoration?

Could the authors clarify how the theoretical guarantees in Section 3.5 apply to the actual adaptive optimization procedure? In particular, can the complementarity assumptions and finite-fitting residuals in Appendix I be empirically characterized? Similarly, does the proposed temporal regularization guarantee any meaningful stability of the reconstructed endpoints when the spatial factors and network parameters are simultaneously updated?

Could the authors provide a more controlled ablation comparing the complete framework against simpler alternatives, such as conventional robust losses or IRLS-based weighting, static Tucker factorization, and deterministic iterative recovery? Such comparisons would help distinguish the benefits of the proposed components from those of a stronger scene-specific optimization procedure.

Why is Diff-Unmix (CVPR 2024) not included in the denoising comparison? Also, why does the MP-HSIR evaluation use Natural_scene.ckpt rather than the dedicated remote-sensing checkpoint provided by its official implementation? Could the authors clarify whether the corresponding dataset partitions and preprocessing procedures are consistent with the original baselines?

Could the authors provide additional experiments under more conventional HSI degradation protocols and across multiple image crops and independently sampled corruptions? The current setup uses a custom complex-noise distribution and dataset-specific optimization parameters. Clarifying how these parameters were selected, and reporting mean and variance over multiple trials, would help establish the robustness and fairness of the results.

Given the substantial computational overhead, how does restoration quality change with the number of diffusion steps and optimization updates? A PSNR–runtime comparison against DDS2M and representative optimization-based methods under matched computational budgets would help determine whether the proposed framework offers a practically meaningful advantage.

3 comments

Thanks.

Response to Q1 & CW 1

We apologize for not clearly highlighting the methodological innovations. Our contribution is not merely a more elaborate likelihood model or additional structural regularization, but a novel uncertainty-aware and trajectory-aware self-supervised diffusion framework. First, to our knowledge, VITTA provides the first theoretically characterized Gaussian-projection mechanism for adapting complex non-Gaussian corruption to self-supervised diffusion sampling, a design which is entirely absent from DDS2M. Second, VITTA pioneers the explicit modeling of structural evolution across reverse diffusion steps, emphasizing crucial temporal dependencies largely ignored by existing methods. These fundamental innovations distinguish VITTA from DDS2M.

Designs target HSI traits: (i) prior-free self-supervision for data scarcity and varying dimensions; (ii) dense-sparse modeling for real HSI noise; and (iii) Temporal Tucker for HSI low-rankness. Yet VITTA is not limited to HSI; we evaluate on HSI simply because it is the most suitable domain. We leave extensions to other tensor data as future work.

Response to Q2 & CW 1

While our model does not explicitly capture spatial correlations, our formulation is actually a more reasonable choice. Its sparse component already captures stripes and dead lines, consistent with standard HSI noise modeling [1], while avoiding specific spatial assumptions ensures greater generality across real HSIs without such patterns.

Zero-centered projection is a deliberate simplification, aligning with prior zero-mean assumptions [2] while ensuring stable and tractable optimization. And nonzero center is not necessary since sampling does not heavily depend on the bias, relying instead on the variance to measure uncertainty.

[1] Hyperspectral Mixed Noise Removal by ℓ₁-Norm-Based Subspace Representation

[2] Adaptive Hyperspectral Mixed Noise Removal

Response to Q3 & CW 1

Guarantees such as finite-fit errors and observation confidence can be empirically assessed. On Chikusei denoising, ρt=1−γt(1−qtc)\rho_t = 1 - \gamma_t(1-q_t^c) averaged 0.9900.990 (ρt<1\rho_t < 1 over all steps); with γt>0\gamma_t > 0, qtc<1q_t^c < 1, showing measurement–regularization synergy. Also, ∥εt−1fit∥/n=1.618±2.947\lVert \varepsilon_{t-1}^{\mathrm{fit}} \rVert/\sqrt{n} = 1.618 \pm 2.947, mean η^t/n=0.0516\widehat\eta_t/\sqrt{n} = 0.0516, and final RMSE fell from 0.47450.4745 to 0.05080.0508, supporting stability analysis. Temporal regularization limits Tucker-core temporal curvature to suppress fluctuations and step-wise overfitting, with network updates handled in Appendix I. Table 5 shows its effectiveness.

Response to Q4 & CW 2

We replaced our variational adaptation as required, obtaining PSNR/SSIM/SAM results of 31.254/0.8375/0.1025 (Huber) and 30.655/0.8118/0.1073 (IRLS), compared with 33.138/0.8911/0.0957 for VITTA, averaged across all three datasets on the denoising task. Table 4 already provided comparisons with static Tucker (Ours w/o Temporal Coupling) and deterministic recovery (Ours w/ DIP). These results show that our gains stem from algorithmic design rather than scene-specific optimization.

Response to Q5 & CW 2

Diff-Unmix is excluded as it relies on augmented samples or external CAVE pretraining, violating our zero-shot setting. MP-HSIR’s remote-sensing ckpt is not used as its training sets overlap test sets (Chikusei and PaviaU), and its fixed 100-band input cannot handle 48-band Houston 2018. For preprocessing, we follow protocols adopted in Hipandas, DDS2M and HLRTF.

Response to Q6 & CW 2

Our noise simulation follows Hipandas, and our model characterizes general corruption rather than fitting Eq. (6). We tested VITTA under degradation protocols of Ours, HLRTF and GNRSR across three crops and three seeds. Due to space limitations, averages of the protocol-wise means and standard deviations are reported (more details are in Appendix L.5), confirming generalizability. VITTA selects parameters via NIQE and Spectral Roughness without GT, which is reasonable and common. Competing methods were tuned to be optimal.

Table 1

TaskMethodPSNR (dB) ↑\uparrowSSIM ↑\uparrowSAM (rad) ↓\downarrow
DenoisingHLRTF30.017±1.3560.7797±0.03450.1217±0.0323
VITTA31.618±1.9350.8512±0.03340.1072±0.0304
CompletionGNRSR29.473±1.6490.7604±0.04770.1511±0.0425
VITTA31.083±1.8590.8333±0.03430.1188±0.0348
PansharpeningHipandas26.678±1.2450.7277±0.04870.2867±0.0843
VITTA28.432±1.8410.7942±0.03430.2082±0.0614

Response to Q7 & CW 2

Restoration quality progressively improves with increasing optimization iterations. Despite runtime overhead, VITTA yields substantial PSNR gains, e.g., surpassing SOTA by 1.24 dB on Chikusei denoising. Due to space limits, updated Appendix L.4 gives the convergence curve and PSNR–runtime plots, showing VITTA still performs best when matching DDS2M in complexity.

Reviewer_qFhTOpenReview ↗

Response to Authors’ Rebuttal

I thank the authors for their detailed rebuttal and additional experiments. The new evaluations under multiple degradation settings address several of my concerns regarding experimental validation. However, my main concern regarding methodological novelty remains. While the rebuttal clarifies the differences from DDS2M, the fundamental technical insight and necessity of the proposed combination of existing techniques are still not sufficiently convincing. Overall, I appreciate the authors' efforts and acknowledge the improved experimental evidence, but I maintain my original rating.

We sincerely thank the reviewer for the thoughtful feedback and careful consideration. We would respectfully like to further clarify that VITTA is not merely a combination of existing techniques, but represents a methodological advance: it extends the applicability of diffusion models to complex non-Gaussian degradations through a principled projection mechanism and provides theoretical guarantees for self-supervised restoration under explicit assumptions, moving beyond purely empirical justification. Just as DDS2M advanced DDRM by introducing self-supervised inference, we believe VITTA takes a further methodological step by introducing a theoretically guaranteed projection paradigm for non-Gaussian inverse problems, shifting self-supervised diffusion restoration from heuristic, Gaussian-dependent designs toward principled and distribution-agnostic solvers. We sincerely appreciate the reviewer's valuable time and constructive feedback.

Reviewer_47FWOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VITTA, which restores a hyperspectral cube at test time. At each reverse step, it fits a bandwise dense precision and a sparse Student-t component, matches the residual to a Gaussian variance, and routes that variance into a singular-vector update. A Temporal Tucker module, with a DIP spatial generator and a cosine core path, estimates the clean endpoint. Contribution 2 and Section 3.2 present the variational bridge as developed here, although Appendix A.1 already cites mixture variational noise and reverse-time precision estimation [Cheng et al., 2023; Cheng and Tan, 2024]. Equation (24) is a conditional state-error recurrence, and Appendix J.2 synthesizes the same dense-plus-sparse family as Eq. (6) at 1 to 1.2 dB SNR.

Critical strengths

The conditioning interface is implementable. Equations (9) and (12) give the variational moments and projected variance, and Eq. (18) states which singular directions trust the measurement. One code path covers denoising, band completion, and pansharpening through HH and the optional PAN term. Table 4 is a real ablation: two-factor noise, Tucker factors, temporal coupling, and DIP versus diffusion are separated, and the averages track Table 1. Band completion is the strongest result: Chikusei reaches 33.148 PSNR and 0.0648 SAM against 31.053 and 0.0784 for HLRTF (Table 2) [Luo et al., 2022]. Pansharpening PSNR is 29.139, 28.398, and 31.393 against 26.934, 27.437, and 29.721 for Hipandas (Table 3) [Xu et al., 2025].

Critical weaknesses
  1. Section 2 and Section 3.4 cite the endpoint loss and singular-vector sampler for DDS2M and DDRM [Miao et al., 2023; Kawar et al., 2022]. Appendix A.1 already places the observation bridge in ScoreDVI and DiffusionVI [Cheng et al., 2023; Cheng and Tan, 2024], then says VITTA develops that component for hyperspectral data. Contribution 2 and Section 3.2 omit those citations. They start from [Andrews and Mallows, 1974] and present variational likelihood plus Gaussian projection as developed here. The HSI-specific delta is the bandwise precision, the sparse weight risr_i^s, and δj,t\delta_{j,t} in Eq. (18). That delta is smaller than in Contribution 2. Section 2 also says the DDS2M endpoint lacks joint spatial-spectral structure, while Appendix A.2 credits DDS2M with spatial-spectral networks.

  2. Contribution 2 claims that Gaussian projection preserves quadratic risk and thereby guarantees faithful spectral-spatial recovery. Proposition 2 only equates mean, covariance, and quadratic risk for a fixed affine map A(e,ξ)=a+Le+CξA(e,\xi)=a+Le+C\xi under a fixed channel law (Eq. (20)). It does not bound SAM. Contribution 1 says recursive endpoint bounds verify progressive convergence. Equation (24) defines Et=(E∥ut−u0∥2)1/2E_t=(\mathbb{E}\|u_t-u_0\|^2)^{1/2}, a diffusion-state norm, and Appendix I.4 obtains ρt<1\rho_t<1 only under an unmeasured margin γt>0\gamma_t>0.

  3. Appendix J.2 draws dense SNR from U(1,1.2)U(1,1.2) dB and then assigns about 30 percent of interior bands to stripes, 30 percent to dead lines, and 30 percent to impulse noise. That process matches Eq. (6), so the largest Table 4 gain is measured on a matched noise model. Every main number uses one central crop, degradation seed 2025, and restoration seed 3407 (Appendix J.1, Appendix K.3). Houston denoising changes π0\pi_0 to (0.995,0.005)(0.995,0.005), and Table 6 changes ranks by dataset. Pavia denoising is then 31.517 versus 31.338 PSNR for GNRSR (Table 1) [Tan et al., 2025]. Table 3 omits DDS2M, which Table 4 lists at 24.567 average pansharpening PSNR.

Questions for authors
  1. Relative to ScoreDVI and DiffusionVI [Cheng et al., 2023; Cheng and Tan, 2024], what is new in Eq. (6) to Eq. (13) beyond bandwise Student-t scales and the plug-in of δj,t\delta_{j,t} into Eq. (18)? Please answer with a line-by-line delta against DDS2M and DDRM as well [Miao et al., 2023; Kawar et al., 2022].

  2. Please report ρt\rho_t, ηt\eta_t, and the fraction of steps with γt>0\gamma_t>0 on one Chikusei denoising run (Eq. (24), Appendix I.4). Should Contribution 2 claim a guarantee of faithful recovery, or only second-order invariance for a fixed affine channel map (Eq. (20))? The abstract should follow that answer.

  3. Were the Table 6 ranks and the Houston π0\pi_0 chosen on the crops in Tables 1 to 3? Please freeze one hyperparameter vector before evaluation, repeat at least three crops and three seeds, and add the higher-SNR Gaussian, stripe, and impulse settings used by HLRTF and GNRSR [Luo et al., 2022; Tan et al., 2025], which are not generated from Eq. (6). A gain that survives that protocol would raise my rating. More scenes under the current matched 1 dB draw would not.

  4. Please add DDS2M to Table 3 on the same pansharpening draw used for Hipandas [Xu et al., 2025]. Please also define spectral roughness and the NIQE band aggregation used in Appendix L.3.

5 comments

Thanks. We emphasize that VITTA offers genuine methodological novelty and consistently achieves superior performance under the required settings.

Response to Q 1 & CW 1

VITTA’s novelty stems from distinct methodological and theoretical foundations beyond merely introducing bandwise Student-tt scales or altering noise parameters. Due to space limits, we clarify the line-by-line delta across key dimensions:

  1. Methodological differences. ScoreDVI jointly optimizes image and noise posteriors via pretrained Gaussian-denoiser priors, making structure–corruption separation sensitive to prior mismatch. DiffusionVI performs successive MAP fusion under a diagonal-Gaussian likelihood, where unmodeled noise correlations may distort observation weights and propagate reconstruction errors. DDRM and DDS2M, meanwhile, rely heavily on Gaussian assumptions, lacking VITTA’s Gaussian projection and trajectory-aware mechanism. Conversely, VITTA explicitly models background noise and severe corruption via Gaussian projection and incorporates Temporal Tucker for spatial–spectral regularization, enabling prior-free, reliability-guided diffusion.

  2. Theoretical differences. VITTA establishes whether complex-noise approximation preserves reconstruction-relevant information and how errors propagate during reverse diffusion, ignored by prior methods.

  3. Practical advantages for HSI. ScoreDVI, DiffusionVI, and DDRM rely on pretrained priors—hard to obtain across diverse HSI spectral dimensions and distributions—while DDS2M lacks modeling of complex HSI corruption. VITTA overcomes both via self-supervised, corruption-aware restoration.

Regarding DDS2M's spatial–spectral modeling, we have revised our previous imprecise statement.

Response to Q 2 & CW 2

On the Chikusei denoising run, ρt\rho_t ranged from 0.985 to 0.999 (mean 0.9902), the normalized empirical η^t/n\hat{\eta}_t/\sqrt{n} from 0.0014 to 0.4928 (mean 0.0516), and γt>0\gamma_t>0 for 100% of steps, providing empirical support for the contraction conditions along this trajectory. For Contribution 2, we agree Eq. (20) establishes second-order invariance for a fixed affine channel map rather than unconditional recovery. Yet Gaussian projection also preserves transition covariance and quadratic restoration risk, while Propositions 3–4 bound and attenuate residual higher-order distributional discrepancy, substantiating reliable reconstruction. We have revised Contribution 2 and the abstract.

Response to Q 3 & CW 3

VITTA selects parameters (e.g., Table 6 ranks, Houston π0\pi_0) on each evaluated test crop via no-reference NIQE and Spectral Roughness without GT. Since varying scenes and degradations require different model capacities and regularization strengths, fixed hyperparameter is not appropriate in self-supervised restoration and thus is not adopted, and our adaptive selection strategy is more reasonable and consistent with prior studies. Competing methods were optimized to be optimal. We evaluated VITTA on the three datasets across three crops, three seeds, and settings used by Ours/HLRTF/GNRSR configurations as requested. Due to space limitations, we report the mean over all runs. VITTA's superiority demonstrates strong robustness and generalizability beyond the original noise setting. More details are in Appendix L.5.

O/H/G = Ours/HLRTF/GNRSR noise; cells: PSNR↑/SSIM↑/SAM↓.

Table 1. Denoising

NoiseHLRTFGNRSRLMELTVVITTA
O31.765/0.8339/0.105631.990/0.8514/0.106530.909/0.8181/0.250132.524/0.8856/0.0907
H30.121/0.7817/0.120529.980/0.7894/0.149129.244/0.7506/0.221730.473/0.8438/0.1124
G28.164/0.7236/0.139129.470/0.7460/0.157829.289/0.7143/0.293131.856/0.8243/0.1185

Table 2. Missing-Band Completion

NoiseHLRTFGNRSRLMELTVVITTA
O30.839/0.8078/0.116830.680/0.8053/0.122829.507/0.7678/0.254832.219/0.8727/0.0949
H29.949/0.7712/0.132529.188/0.7647/0.158328.787/0.7279/0.247430.019/0.8231/0.1284
G27.976/0.7076/0.151828.551/0.7112/0.172228.280/0.6702/0.333131.012/0.8041/0.1331

Table 3. Noisy Pansharpening

NoiseSERT+PLRDMP-HSIR+PLRDHipandasVITTA
O27.794/0.7703/0.194126.714/0.7231/0.209027.554/0.7794/0.237328.761/0.8097/0.1831
H26.468/0.7254/0.308924.873/0.6565/0.365624.659/0.6441/0.373227.465/0.7738/0.2409
G29.037/0.7981/0.227429.019/0.7838/0.221827.821/0.7597/0.249529.070/0.7992/0.2007

Response to Q 4 & CW 3

We have added DDS2M to Table 3, obtaining results of 24.771/0.6376/0.2431, 22.764/0.5426/0.4844, and 26.166/0.6366/0.5060 (PSNR/SSIM/SAM) on Chikusei, PaviaUniversity, and Houston 2018, while VITTA still achieves the best performance. Spectral roughness and NIQE procedure have been clarified in updated Appendix L.3.

Thanks for the detailed response by the authors. This response has addressed most of my concerns. However, I found that some very relevant works should be surveyed and included in the revised paper, including:

[1]. Yu, Dabing, and Zheng Gao. "VolFormer: Explore more comprehensive cube interaction for hyperspectral image restoration and beyond." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025. (Cube-interaction modeling for HSI restoration aligns with VITTA’s spatial–spectral–temporal structure prior.)

[2]. Yu, Dabing, et al. "Dstrans: Dual-stream transformer for hyperspectral image restoration." 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2023. (Dual-stream Transformer HSI restoration is a key recent learning-based spatial–spectral competitor.)

[3]. Yang, Xingxing, Jie Chen, and Zaifeng Yang. "Hyperspectral image reconstruction via combinatorial embedding of cross-channel spatio-spectral clues." Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 7. 2024. (Cross-channel spatio-spectral embedding for HSI reconstruction is a strong supervised spatial–spectral baseline.)

[4]. Li, Miaoyu, et al. "Latent diffusion enhanced rectangle transformer for hyperspectral image restoration." IEEE Transactions on Pattern Analysis and Machine Intelligence 47.1 (2024): 549-564. (discuss latent diffusion-based for HSI restoration)

[5]. Yang, Xingxing, Jie Chen, and Zaifeng Yang. "Unraveling Metameric Dilemma for Spectral Reconstruction: A High-Fidelity Approach via Semi-Supervised Learning." Advances in Neural Information Processing Systems 38 (2026): 134423-134446. (Semi-supervised spectral reconstruction addressing metamerism is closely related to spectrally faithful HSI recovery.)

If all these concerns could be addressed, I would like to raise my evaluation.

Thank you for your valuable suggestions. We have incorporated all five recommended references into the revised manuscript and added relevant discussions. We sincerely appreciate your constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.