acceptodds
Under review as a conference paper at ICLR 2027

Beyond Feature Modulation: Physics-Conditioned State Evolution for Hyperspectral Compressive Imaging

Abstract

Physical acquisition information provides valuable guidance for reconstructing hyperspectral images from compressed measurements. Existing methods have incorporated acquisition information through mask-guided feature modulation or adaptive optimization, yet such strategies mainly affect current representations rather than explicitly governing state evolution that carries accumulated information into subsequent inference. Motivated by this, we propose Physics-Conditioned State Evolution (2State), a framework that uses state space models to extend physical guidance from feature modulation to state evolution. Specifically, we design a deep unfolding network that couples explicit physical modeling and implicit prior inference through a shared effective coding operator. For explicit physical modeling, a bounded correction of the nominal mask yields an effective operator that guides data-consistency updates. For implicit representation, the same corrected mask conditions selective state space models within the proximal network to regulate information injection, retention, and readout throughout the entire state evolution process. The evolving states enable acquisition-dependent information to be progressively accumulated, updated, and reused during proximal inference. Extensive experiments on hyperspectral compressive reconstruction with varying spectral dimensionalities demonstrate the effectiveness of the proposed framework. Ablation studies and state evolution analyses further validate the proposed guidance mechanism and characterize its role in shaping inference dynamics.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Hyperspectral compressive imaging, physics-conditioned state evolution, state space models, mask-guided fusion, deep unfolding

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_HxCyOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Physics-Conditioned State Evolution (Φ2State) for hyperspectral compressive reconstruction. The method uses a learnable coding mask in a deep unfolding network. The corrected mask is used for both data consistency and state evolution in Mamba. The physical information is introduced into the state parameters to control information updating, retention, and readout. The authors evaluate the proposed method on four hyperspectral datasets and conduct several ablation studies to validate its effectiveness.

Critical strengths
  1. Integration of physical information into state space modeling.

The idea of using physical acquisition information to guide the internal state evolution of Mamba is interesting. Instead of only using the coding mask for feature modulation, the proposed method introduces physical information into the state parameters to control information updating, retention, and readout. In addition, the shared effective coding mask provides a reasonable connection between data consistency and learned reconstruction. This design may help the network better use physical information during the reconstruction process.

  1. Evaluation on different hyperspectral datasets.

The authors evaluate the proposed method on four hyperspectral datasets with different spatial and spectral characteristics, covering 31 to 144 spectral bands. The experiments include comparisons with several existing methods and ablation studies on the main components. In particular, the ablation studies investigate the effects of the learnable coding mask, different state evolution strategies, and physical modulation of state parameters. These experiments provide useful information for understanding the proposed method and its performance.

Critical weaknesses
  1. Limited performance improvements and insufficient justification for the proposed state evolution mechanism.

The performance improvements from the main components are relatively small. In Table 3, introducing State Evolution improves PSNR by only 0.18 dB, while Stage Refinement provides an additional 0.10 dB. In Table 4, the proposed Physics-conditioned State outperforms Cross Attention by only 0.15 dB, with the same SSIM of 0.9613. In Table 5, the full modulation of B, C, and Δ achieves only 0.09 dB improvement over modulating C alone. Some combinations of these parameters even perform worse than using a single parameter. These results make it difficult to understand why the proposed state evolution mechanism is necessary compared with simpler alternatives.

  1. Insufficient validation of physical mask correction and missing computational efficiency comparisons.

The proposed method introduces a learnable effective coding mask to improve the physical acquisition model. However, the paper does not clearly demonstrate whether the estimated mask actually corrects physical acquisition errors or simply acts as additional learnable parameters to improve reconstruction performance. There is no quantitative evaluation or visualization showing how the coding mask is corrected and whether the correction is physically meaningful.

In addition, the authors mention GFLOPs and model parameters (Params) as evaluation metrics, but the corresponding results are not reported. Since the proposed method introduces additional modules for mask correction and physics-conditioned state evolution, computational efficiency comparisons are necessary. Without these results, it is difficult to determine whether the relatively small reconstruction improvements justify the additional model complexity.

Questions for authors

1.How can the authors demonstrate that the learned effective coding mask provides meaningful physical correction rather than simply improving reconstruction through additional trainable parameters? Could the authors provide quantitative results or visualizations to show how the original coding mask is corrected?

  1. Necessity of state evolution: The improvements from the proposed state evolution mechanism are relatively small in Tables 3–5. Could the authors further explain why this mechanism is necessary compared with simpler methods such as feature conditioning or cross-attention? Are these improvements consistent across multiple training runs?

  2. Computational complexity: The paper mentions GFLOPs and Params as evaluation metrics, but the results are missing. Could the authors provide GFLOPs, parameter counts, and inference times for the proposed method and competing methods? This would help evaluate whether the performance improvements justify the additional computational cost.

  3. Stage-wise mask estimation: In Equation (5), the effective coding mask is estimated from the measurement and the original mask, without using the current reconstruction result. However, Section 3.2 states that the mask estimation adapts to the evolving reconstruction state. Could the authors clarify how the estimated mask is adapted to the reconstruction state across different unfolding stages?

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

This paper proposes a physical conditional state evolution framework for hyperspectral compressed imaging reconstruction. This method extends physical guidance from feature modulation to state evolution through a state-space model. Specifically, the authors couple explicit physical modeling with implicit prior reasoning by sharing effective coding operators. Bounded modifications to the nominal mask are used to obtain operators that guide data consistency updates. Information control is achieved by constraining the state-space model with the modified, identical mask, enabling the gradual accumulation, updating, and reuse of information dependent on the acquired data during proximal inference. Experiments were conducted on four hyperspectral datasets, achieving competitive reconstruction performance. The effectiveness of each module was analyzed through ablation experiments.

Critical strengths

1.This paper presents a well-designed methodology. Based on a clear task motivation, the authors propose a framework that effectively combines physical information with a state-space model. This is achieved by coupling explicit physical modeling with implicit prior reasoning through shared effective coding operators. Introducing shared coding conditions into the dynamic reasoning process allows for the gradual accumulation and reuse of reconstructed evidence, establishing an effective link between physical reconstruction and learning-based prior reasoning. This provides a valuable research approach for reference.

2.The experimental evaluation in this paper is quite comprehensive, covering hyperspectral datasets with four different spectral dimensions, and comparing with several representative state-of-the-art reconstruction methods, achieving competitive results. The ablation experiments also thoroughly decompose and validate the main design, providing corresponding evidence to demonstrate the necessity and rationality of the design.

Critical weaknesses

1.This paper does not sufficiently verify the physical significance of learnable coded mask correction. The paper constructs an effective coding operator through learnable mask correction, but the existing experiments mainly reflect the improvement in reconstruction accuracy, and have not fully demonstrated whether the corrected coding operator can more accurately reflect the actual measurement process.

2.The analysis of model computational efficiency in this paper is insufficient. The paper mentions efficiency indicators such as model parameters, but does not provide corresponding comparative analysis. The paper does not report the corresponding computational overhead, making it impossible to comprehensively evaluate the actual benefits of the proposed method.

Questions for authors

1.The paper constructs an effective sensing operator by learning a corrected coding mask. Could you further explain the correction mechanism? Especially when the original coding mask is known, what acquisition errors does the corrected operator primarily compensate for?

2.Table 4 in the paper compares different physical condition state evolution strategies. Could you further explain whether these schemes use the same network backbone, input, and training settings? Are the parameter sizes of the related schemes comparable?

3.Section 4.1 of the paper mentions using GFLOPs and model parameter count as computational efficiency evaluation metrics, but the paper does not provide relevant results. Could you supplement with relevant experiments and their result comparisons?

  1. Why not use the simulated and real datasets used by most baseline methods? There's a lack of comparisons with state-of-the-art methods like LADE-DUN, DAUHST, etc. Also, why is the visualization of the cave dataset's partitions placed in the main text, unless it's your own dataset? The main text contains too much unnecessary content, lacks experiments, and has too few visual comparisons to easily determine the method's superiority. Resolving these questions would change the score.
Reviewer_VaNCOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies hyperspectral image reconstruction from CASSI measurements. It aims to improve the use of physical acquisition information during reconstruction. The proposed method combines deep unfolding with a learnable correction of the sensing mask. The corrected mask is used for data consistency and is also injected into the selective parameters of Mamba. Experiments on multiple hyperspectral datasets show PSNR improvements over the compared methods.

Critical strengths

The paper studies a relevant problem in hyperspectral compressive imaging. The physical formulation and unfolding process follow established reconstruction methods.

The main design is straightforward and easy to follow.

The proposed method achieves the highest average PSNR on all four datasets.

The results show some positive contributions from the proposed components.

Critical weaknesses

The main motivation is not sufficiently justified. The paper argues that existing methods lack an explicit mechanism to preserve and reuse acquisition-dependent information. However, Mamba already maintains hidden states through recurrent updates. The proposed method mainly adds physical embeddings to the parameter generation of B, C, and Δ. It remains unclear what new state modeling capability is introduced.

The proposed mask correction lacks sufficient justification. The paper learns a correction to the nominal sensing mask. However, the physical mask is already known in the simulated acquisition process. It is unclear why a corrected mask should provide a more accurate physical model. Besides, to my understanding Eq. (5) only takes the measurement and nominal mask as inputs. The correction therefore does not directly depend on the evolving reconstruction state.

The performance improvements are limited, and the comparison protocol is insufficiently detailed. It seems that the proposed method achieves the best SSIM on only one of the datasets. The paper also mentions GFLOPs and parameter counts but does not report these results. Without matched computational costs, training settings, and repeated runs, it is difficult to assess the significance of these improvements. Comparisons with closely related physics-guided Mamba methods are also missing.

The ablation studies do not fully support the proposed state evolution mechanism. The experiments mainly show improvements from adding physical information to Mamba parameters. However, these results do not establish how the hidden states preserve and reuse acquisition information. No direct state evolution analysis is provided. Some results in Table 5 also contradict the claim that combining more conditioned parameters consistently improves performance. For example, conditioning B and Δ performs worse than conditioning B alone.

The lack of real CASSI experiments and mask perturbation analysis further limits the evaluation.

Questions for authors

How does the proposed state evolution differ from the standard recurrent mechanism in Mamba? Are hidden states explicitly transferred between unfolding stages? If so, could the authors provide direct evidence of this process?

Why is the learnable correction of the nominal sensing mask necessary? Could the authors evaluate this component under controlled mask errors and compare the corrected operator with the true sensing operator?

Could the authors provide parameter counts, GFLOPs, and matched training settings for the compared methods? A direct comparison with recent physics-guided Mamba methods would also help clarify the contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.