acceptodds
Withdrawn from ICLR 2027

Learning How Factors Interact: Conditioned Tensor Products for Unsupervised Hyperspectral Super-Resolution

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Hyperspectral image super-resolution (HSI-SR) aims to reconstruct a high-spatial-resolution hyperspectral image from a low-resolution hyperspectral observation and a registered high-resolution multispectral image (HR-MSI). Existing unsupervised factorized priors typically employ globally fixed factor interactions and discrete band-wise spectral parameterizations, limiting adaptation to spatially varying guidance and ordered spectral variation. To address these limitations, we propose CPF-RFJ, a unified test-time framework that jointly adapts factor interactions and spectral representations. We first propose the Conditioned Tensor Product (CTP), which allows auxiliary observations to select the multiplication law between latent factors, thereby shifting conditioning from feature representations to the algebraic interaction itself. Building upon CTP, the Conditioned Product Factorization (CPF) enables the local HR-MSI observation at each spatial location to govern spectral–spatial factor coupling. We further develop a Pole-Free Rational Fractional-Jacobi Kolmogorov–Arnold Network (PF-RFJ-KAN), which models the spectral factor as a coordinate-defined function through signed fractional spectral reparameterization, an adaptive Jacobi basis family, and a structurally pole-free rational representation. The complete CPF-RFJ model is optimized directly from the observations without external training data or target HR-HSI supervision. Extensive experiments demonstrate state-of-the-art spatial–spectral reconstruction performance against representative training-data-free methods.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Hyperspectral Image Super-Resolution, Conditioned Tensor Product, Tensor Factorization, Deep Image Prior, Kolmogorov–Arnold Network

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_NSdfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates hyperspectral image super-resolution without external training data, reconstructing a HR-HSI by fusing a LR-HSI with a HR-MSI. The authors propose the CTP and its positive diagonal instantiation, CPF, which generate interaction weights for latent factor blocks using local MSI observations. Furthermore, they introduce the PF-RFJ-KAN to construct residual corrections for spectral factors via learnable fractional coordinates, Jacobi bases, and rational functions with strictly positive denominators. The complete model is optimized at test time through a dual-observation consistency loss. The primary algebraic properties, the pole-free guarantee, and the proofs of regularity, gradient bounds, and density in the appendix are mathematically sound. Table 1 shows that this configuration achieves the best results on 22 and the second-best on 2 out of 24 metric comparisons across three simulated dataset crops and two spatial degradation scales. The main concerns lie in the novelty positioning of the conditioned interaction, the empirical attribution of performance gains to core components, and result consistency across different tables and figures.

Critical strengths
  1. The task holds practical value, and the overall methodology is well-motivated. Local MSI patches provide structural guidance for high-resolution spatial details, spectral-coordinate residuals complement the deep prior factors, and the dual-observation consistency objective aligns naturally with the fusion setting without external training data. Table 3 explicitly uses the observation loss on clean inputs for checkpoint selection, a procedure that strictly adheres to the supervision-free target HR-HSI setting.
  2. The primary mathematical derivations are sound, and the pole-free design serves a clear purpose. Bilinearity, associativity, and norm bounds under fixed conditioning hold as stated; the denominator is formulated as 1 plus squared basis terms with positive coefficients, guaranteeing a lower bound of 1. The proofs in the appendix regarding the Jacobi recurrence, fractional-coordinate regularity, parameter gradient bounds, and approximation density as the degree increases are also valid. Notably, the authors explicitly clarify that associativity requires an invariant conditioning state and qualify that the density result does not imply optimization advantages for finite architectures. 3.The complete model demonstrates competitive performance under the reported settings. The 22/24 best-ranking statistics in Table 1 are consistent with the PSNR gains reported in the main text, and the slightly inferior SSIM and SAM compared to UCL on PU (×8\times 8) are transparently acknowledged. Moreover, the relative MAE improvements in Figure 5 and the MSI observation error reduction ratios in Table 4 are accurately calculated.
  3. The narrative is well-structured, and key implementation details are transparently disclosed. The methodology unfolds logically across conditioned interaction, spectral residual correction, and observation-consistent optimization, accompanied by specifications of initialization schemes, optimization schedules, parameter constraints, software versions, runtime, and GPU memory footprint. In the real-data experiments, the authors also explicitly note that MSI observation residuals should not be conflated with ground-truth HR-HSI reconstruction errors.
Critical weaknesses
  1. The positive diagonal gating in Eq. (6) and Eq.(15) can be absorbed into the spatial factor, reformulating the reconstruction as a standard two-factor inner product. Meanwhile, the normalization in Eq. (61) alters the effective factor constraints once the gating is absorbed. The authors should clarify whether the performance gains stem from the conditioned inductive bias, variations in the constrained per-pixel magnitude, or optimization dynamics, in order to substantiate the core novelty claim of "learning a multiplication law."
  2. The PSNR values for PaviaC (×16\times 16) in Table 1 and Table 2 are 45.6395 and 45.6851 dB, respectively—a discrepancy larger than the 0.0092 dB gain of CPF over the matrix product reported in Table 2. Similarly, the red dots representing default hyperparameter settings for the same scene in Figure 7 also exhibit discrepancies. The authors should specify the exact configurations, random seeds, and checkpoint selection sources to account for these inconsistencies.
  3. Table 3 reports only a single fitting run per scene–scale pair, making it impossible to assess whether the marginal interaction gains are statistically stable. Moreover, Figure 5 lacks a baseline without spectral residuals and does not isolate the individual effects of the spectral sub-components. I recommend completing a 2x2 ablation matrix (with/without gating ×\times with/without spectral residual) and reporting results across a small number of independent random seeds alongside model capacity and optimization budget specifications for key comparisons.
  4. While Theorem 3.1 and Proposition B.4 are mathematically correct, their bounds depend on the norms of the learned coefficients or residuals; thus, a fixed λ\lambda alone does not guarantee a bounded correction magnitude after optimization. Furthermore, varying the Jacobi parameters does not expand the numerator polynomial space of a fixed degree. In addition, since continuous coordinates are applied solely to the spectral residual branch, claims regarding "controlled correction," "function-space adaptation," and continuous parameterization of the full spectral factor should be aligned with the actual mathematical construction.
  5. Fig. 4 and Table 4 evaluate different sets of baseline methods, yet the configurations of the newly added methods and the rationale behind this discrepancy are left unexplained. The real-data experiments lack sufficient details regarding image registration, band selection/processing, normalization, and degradation operators; for the simulated data, the exact crop coordinates, band indices, SRFs, and complete generator architectures are also omitted. Finally, the exact relationship among the datasets denoted as PaviaC, PC, and Original ROSIS needs to be explicitly clarified.
Questions for authors
  1. Since the positive diagonal CPF in Eqs. (6) and (15) can be equivalently expressed as applying a conditioned scaling to the spatial factor, how do the authors delineate its added value over conventional conditioned factor gating in light of the normalization in Eq. (61)? Please specify the effective constraints on the spatial factors and clarify whether the performance gains stem primarily from a conditioned inductive bias, variations in the constrained per-pixel magnitude, or underlying optimization dynamics.
  2. Why do the PaviaC PSNR values of the full model differ between Table 1 and Table 2? Do the three default parameter points for the same scene in Figure 7 correspond to the identical remaining configurations and degradation scales? Please detail the data cropping, hyperparameter configurations, random seeds, and checkpoint selection sources for these results, and account for the observed discrepancies.
  3. Building upon the existing operator ablations in Table 2, could the authors supplement a full 2x2 comparison matrix (with/without conditioned gating ×\times with/without spectral residual), and report the mean and standard deviation over a small number of independent random seeds for key marginal gains? In addition to the already stated backbone and optimization settings, please provide the latent dimensions, parameter counts, and actual fitting budgets for each alternative module and operator.
  4. Why do Figure 4 and Table 4 evaluate different sets of baseline methods? Please clarify the references and configurations for CSTF and LFCTNN, as well as the acquisition pipeline for real-data registration, band processing, normalization, and the spectral response/spatial degradation operators. As a supplementary evaluation, could the authors report the observation consistency error at the LR-HSI end to simultaneously assess data fitting across both modalities?
  5. What exact data files, spatial crops, and retained spectral bands correspond to PaviaC, PC, and the Original ROSIS, and what are their precise interrelationships? What is the rationale behind selecting the default hyperparameters, and are they shared universally across all scenes? Furthermore, please explicitly specify the structural novelties of PF-RFJ compared to the cited fKAN and rKAN, particularly the respective roles of the signed fractional coordinates and the positive squared denominators.
Reviewer_fdFAOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents CPF-RFJ, an unsupervised hyperspectral image super-resolution framework that integrates Conditioned Tensor Products (CTP), Conditioned Product Factorization (CPF), and a Pole-Free Rational Fractional-Jacobi Kolmogorov–Arnold Network (PF-RFJ-KAN). The proposed method uses high-resolution multispectral observations to adaptively control spectral-spatial factor interactions, while introducing a continuous coordinate-based representation for spectral factors. The framework is optimized directly from the observed LR-HSI and HR-MSI without external training data. Experiments on three datasets under different spatial degradation settings demonstrate competitive reconstruction performance against existing unsupervised methods.

Critical strengths

Relevant research problem. The paper addresses unsupervised hyperspectral image super-resolution, which is important for practical remote sensing applications where paired training data are difficult to obtain.

Promising reconstruction performance. The proposed method achieves strong quantitative results on the evaluated benchmarks, particularly under ×8 and ×16 spatial degradations. The integration of spectral-spatial factorization and observation consistency provides a potentially useful framework for hyperspectral reconstruction.

Critical weaknesses

Insufficient fundamental novelty of the proposed conditioned tensor product. Although CTP is presented as a new algebraic formulation, its practical implementation essentially reduces to observation-dependent weighting of latent factor interactions. The conceptual distinction from existing dynamic weighting, gated fusion, and adaptive tensor factorization mechanisms is not sufficiently established. Moreover, the theoretical properties are largely inherited from invertible transform-induced tensor products, making the claimed theoretical contribution less convincing.

Limited evidence supporting the proposed architectural complexity and generalizability. The framework combines multiple sophisticated components, including conditioned tensor factorization, fractional spectral reparameterization, adaptive Jacobi bases, and rational KAN representations. However, the experimental evidence does not sufficiently justify the necessity of this complexity. In particular, Table 2 shows that CPF improves PSNR over conventional matrix multiplication by only approximately 0.17 dB and 0.01 dB under ×8 and ×16 degradations, respectively. Furthermore, quantitative evaluations are concentrated on three closely related Pavia datasets, leaving the generalization capability across different sensors, spectral characteristics, and acquisition conditions insufficiently established.

Questions for authors

What is the fundamental difference between the proposed Conditioned Tensor Product and existing dynamic weighting, gating, or adaptive tensor factorization mechanisms? Can the authors clarify the unique contribution beyond observation-dependent factor weighting?

The theoretical properties of CTP, including bilinearity, associativity, and stability, appear to follow naturally from existing invertible transform-induced tensor products. What new theoretical insights are established beyond these known properties?

Table 2 shows only marginal improvements from CPF over conventional matrix multiplication, particularly under ×16 degradation. How do the authors justify the necessity of introducing a new conditioned tensor algebra given such limited empirical gains?

What are the individual contributions of fractional spectral reparameterization, adaptive Jacobi bases, and pole-free rational representations? More comprehensive component-wise ablation studies are needed to demonstrate whether each design is necessary.

How does the proposed approach compare with simpler conditional factorization or lightweight spectral modeling alternatives under matched parameter counts and computational budgets?

The quantitative experiments mainly focus on Pavia datasets. How well does the method generalize to hyperspectral datasets acquired by different sensors with different spectral resolutions, noise characteristics, and degradation processes?

How sensitive is the proposed framework to inaccurate spectral response functions, spatial misregistration, and unknown degradation operators in practical remote sensing scenarios?

Can the authors provide additional evidence of optimization stability and computational efficiency, including variability across random initializations and fair runtime comparisons with existing unsupervised methods?

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

This manuscript proposes a framework for hyperspectral image super-resolution with two components: a conditioned tensor product a continuous spectral representation based on classic Kolmogorov–Arnold networks.

Critical strengths

The paper is well organised.

Critical weaknesses

First the formulation is ambitious and mathematically elaborate. In its present form, however, the manuscript does not show that this complexity is needed.

Second concern is on the theoretical basis of the conditioned tensor product. The method assumes that the multispectral observation contains enough information to define the multiplication rule between latent factors. Yet local multispectral measurements carry only coarse spectral information. It is therefore not evident that they can determine the underlying coupling between the spectral and spatial modes.

Moreover, letting the product operator vary from pixel to pixel makes the model more expressive, but this may come at the expense of the structural constraints and identifiability guarantees that make classical tensor decompositions effective in the first place. Recent work on generalised tensor formulations for this problem (TPAMI 2025) has analysed recovery conditions explicitly, and a comparable analysis is missing here. I would encourage the authors to establish whether the CTP retains meaningful algebraic properties and whether its additional freedom improves recoverability rather than simply adding trainable parameters.

Observation-guided deep priors combined with KAN-based implicit guidance, SFIGNet, and tensor–tensor product interactions, TPTransformer, have all been proposed for hyperspectral–multispectral fusion.

The same question of necessity arises for the spectral representation. Continuous spectral modelling and KAN-based reconstruction have both been explored before, for example in SFIGNet and in a KAN-GAN approach. The manuscript, however, does not explain why each of its four ingredients is needed. Since real spectra often contain narrow absorption bands and abrupt transitions, it is far from obvious that the proposed basis family is better suited to such signals than simpler choices.

Questions for authors

See weakness part

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

Related papers

Discussion (0)

Sign in to comment.