acceptodds
Under review as a conference paper at ICLR 2027

PDD-Fusion: Physics-Guided Degradation-Decoupled Learning for Low-Light Infrared-Visible Image Fusion

Abstract

Low-light infrared–visible image fusion is challenged by heterogeneous degradation effects across modalities, spatial regions, and frequency components. Visible low-frequency information can suffer from poor exposure, high-frequency responses may mix useful structures with noise, and infrared observations may exhibit weak local thermal variations. We propose PDD-Fusion, a Physics-guided Degradation-Decoupled framework that explicitly models degradation-related reliability and assigns different degradation factors to specialized fusion operations. Specifically, seven imaging-inspired priors are organized into exposure, high-frequency reliability, and infrared thermal groups, and encoded into spatial and global degradation prompts. Guided by these representations, PDD-Fusion performs high-frequency reliability modulation, low-frequency exposure-aware fusion, and gated infrared residual compensation while retaining direct modality paths for reconstruction. Trained only on LLVIP, PDD-Fusion achieves leading performance on multiple fusion metrics and generalizes effectively to MFD and MSRS without fine-tuning, demonstrating the effectiveness of degradation-specific fusion under low-light conditions.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Infrared-visible image fusion, Low-light image fusion, Degradation decoupling, Physics-guided priors。

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_c7tYOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PDD-Fusion, a physics-guided degradation-decoupled framework for low-light infrared-visible image fusion. It constructs seven imaging-inspired priors to characterize visible exposure, high-frequency reliability, and infrared thermal properties. These priors are encoded into spatial and global prompts to guide three specialized fusion branches. Experiments on LLVIP, M3FD, and MSRS demonstrate strong fusion performance and promising cross-dataset generalization without fine-tuning.

Critical strengths
  1. The paper introduces a coherent degradation-aware fusion strategy that explicitly associates heterogeneous degradation cues with specialized frequency-aware operations. This design provides an interpretable and practically meaningful alternative to conventional feature aggregation, forming the main methodological contribution that supports my positive assessment.
  2. The method demonstrates strong quantitative performance across three public datasets, with promising cross-dataset generalization without fine-tuning. Combined with its compact architecture of only 0.90M parameters and reported inference latency of 9.68 ms, these results provide compelling evidence of practical effectiveness.
Critical weaknesses
  1. The methodological novelty could be more clearly distinguished from existing degradation-aware and frequency-based fusion methods. Although the proposed correspondence between degradation cues and specialized fusion operations is meaningful, a more explicit comparison with closely related approaches would help better establish the contribution.
  2. The evaluation relies on randomly sampled subsets from the three datasets, and the reported results are primarily based on conventional fusion metrics. Providing more details about subset selection and, if feasible, additional evaluations on larger or repeated test subsets would strengthen the reliability of the empirical conclusions.
Questions for authors
  1. Could the authors clarify the key methodological differences between PDD-Fusion and existing degradation-aware or frequency-based fusion methods, particularly the novelty of the specialized fusion operations?
  2. Could the authors explain the discrepancy between the spatial-prompt ablation results in Tables 2 and 4? Were the corresponding experimental configurations identical?
  3. Could the authors clarify how the evaluation subsets were selected and whether all compared methods used consistent preprocessing and evaluation protocols?
8 comments
AuthorsOpenReview ↗

Methodological Differences, Spatial-Prompt Ablation, and Evaluation Protocol

We thank the reviewer for the careful reading and constructive suggestions. The main additions in the revised manuscript are marked in blue.

Q1. What distinguishes the specialized fusion operations?

CDDFuse uses feature correlation to encourage shared low-frequency representations and modality-specific high-frequency representations. PhyFusion uses radiative-transfer and Retinex priors with physical consistency to guide cross-modal fusion. In PDD-Fusion, different degradation cues control different operations during fusion (Sec. 3.2). At the same image location, high relative visible brightness directly reduces the visible contribution to the low-frequency mixture, while a separate gate regulates visible high-frequency detail using visible reliability and infrared directional structure. Visible darkness and local infrared flatness jointly scale a Laplacian correction that enhances existing local infrared contrast.

Section 4.4 and Fig. 5(b) illustrate the exposure operation on one LLVIP scene. Under a 5×5\times gain applied to the visible input, the mean low-frequency visible weight in the glare region fell from 0.059 to 0.016. Table 1 and Fig. 3 provide quantitative and visual comparisons with CDDFuse and PhyFusion.

Q2. Spatial-prompt ablation in Tables 2 and 4

“Module 3” (Table 2) and “Spatial Prompt” (Table 4) use the same configuration, with the degradation priors, specialized branches, and spatial prompt enabled and the global prompt disabled. A transcription error in supplementary Table 4 caused the discrepancy. Both tables now report the same results for this configuration.

Q3. Subset selection and evaluation protocol

For the comparisons in Table 1, we randomly selected 300 image pairs from the official LLVIP test split, 300 from MSRS, and 200 from M$^3$FD. Within each dataset, every method received exactly the same infrared–visible image pairs. We scored all fused outputs using the same evaluation pipeline and the same implementations of EN, MI, QAB/FQ_{AB/F}, VIF, and QG. The LLVIP-trained PDD-Fusion model was evaluated directly on MSRS and M$^3$FD without fine-tuning.

Overall, the revision connects the fusion design to measured branch responses and documents cross-dataset transfer in terms of image-fusion quality. Larger or repeated test subsets would help assess sensitivity to sample selection. Downstream task performance remains open.

Reviewer_c7tYOpenReview ↗

Response to the Authors' Rebuttal

  • To my understanding, CDDFuse is a general image fusion method and does not specifically address image degradation. Therefore, it may not be appropriate to include it as a comparison method for degradation-aware image fusion. In addition, I would like to better understand how the authors handle the joint degradation and fusion problem, and what advantages their approach offers over existing fusion methods designed to address image degradation. This aspect should be further clarified. The current explanation mainly focuses on describing the proposed method itself, without sufficiently highlighting its advantages over existing approaches.

  • Moreover, although the authors have added comparison methods, the corresponding explanations and descriptions have not been updated in the main text. For example, the introduction of comparison methods at the beginning of Section 4.2 remains unchanged.

  • Furthermore, in the revised manuscript, the fusion results of CDDFuse in Fig. 3 appear to preserve the human regions better. I would like to know how the authors interpret this observation.

Thank you for taking the time to review our revision and for sharing your follow-up questions.

Q1. Relation to general-purpose and degradation-aware fusion methods.

We include CDDFuse as a general-purpose baseline because both methods separate and fuse low- and high-frequency information. PDD-Fusion addresses an additional question: when the visible image is degraded, which information should still contribute to the fused result? It uses local exposure to adjust the balance between visible and infrared intensity, and estimates detail reliability separately to decide how much visible texture to retain.

DIVFusion separates scene content from illumination, while LENFusion uses feedback to couple brightness adjustment with fusion. The advantage of PDD-Fusion is that it can make different decisions within the same scene. For example, it can rely more on infrared information in a deep shadow, reduce the influence of visible intensity near a headlight, and still retain useful visible texture in a nearby well-exposed region. It can also preserve useful brightness in a noisy region while reducing the influence of unreliable fine detail. On LLVIP, PDD-Fusion achieves a QAB/FQ_{AB/F} of 0.526, compared with 0.505 for CDDFuse, 0.396 for DIVFusion, and 0.393 for LENFusion.

Q2. Description of the comparison methods in Sec. 4.2.

The opening paragraph of Sec. 4.2 (p. 8, lines 410–414) in the revised manuscript lists the comparison methods as follows.

We compare PDD-Fusion with nine representative fusion methods: LRRNet, DDFM, U2Fusion, CDDFuse, Text-IF, LENFusion, DIVFusion, ReCoFuse, and PhyFusion. Table 1 reports the quantitative results on LLVIP, M³FD, and MSRS.

Both added methods are cited in this paragraph.

Q3. Pedestrian visibility in Fig. 3.

We agree that CDDFuse makes the pedestrian stand out more clearly against the background in the second LLVIP example. In the same scene, the road surface and crosswalk markings are easier to see in the PDD-Fusion result. In the last MSRS example, PDD-Fusion also shows the roadway and roadside vegetation more clearly. Across the scenes shown in Fig. 3, PDD-Fusion gives a clearer and more consistent view of the surrounding scene under different lighting conditions.

Reviewer_3iGPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a Physics-guided Degradation-Decoupled framework that explicitly models degradation-related reliability and assigns different degradation factors to specialized fusion operations. However, most of the seven proposed priors are empirically derived and lack theoretical support.

Critical strengths

1.This paper decomposes infrared and visible images into the frequency domain for processing, which is a sound and insightful approach. 2.The paper is well-organized and easy to follow.

Critical weaknesses

1.The ablation study is relatively simplistic and lacks visual comparisons. 2.Most of the seven proposed priors are empirically based, and the method involves too many hyperparameters without stated principles for setting them.

Questions for authors

1.The seven source-derived priors constitute one of the main contributions. However, prior modulation operations of this kind are common in image restoration, and the approach appears to rely largely on empirical design, offering limited innovation. 2.The authors decompose the source images into high- and low-frequency components using a simple Gaussian filter, which is rather coarse. Why not consider other frequency-domain transformation tools? 3.The paper contains many manually set thresholds or hyperparameters. How were these values determined? 4.Moreover, none of the ablation experiments demonstrate the effectiveness from a visual perspective.

6 comments
AuthorsOpenReview ↗

We explain our design choices and present hyperparameter sensitivity tests and qualitative ablations.

We thank the reviewer for the careful reading and constructive suggestions. We have clarified the design rationale, documented the parameter settings, and added qualitative ablations in the revised manuscript.

Q1. Physical motivation and methodological contribution

Our contribution is to use the seven imaging-inspired cues to control three specific fusion operations (Secs. 3.1–3.2 and Appendix A).

  • Visible exposure. We use low-pass visible luminance as an image-relative exposure cue. The brightness, darkness, and high-brightness maps guide the spatial mixing of visible and infrared low-frequency features. In particular, the high-brightness cue directly reduces the visible weight in this mixture (Eqs. 15–17).

  • High-frequency reliability. Under the local image model Y=ℓR+ηY=\ell R+\eta, where ℓ\ell is illumination, RR is reflectance, and η\eta is noise, slowly varying illumination gives H(Y)≈ℓH(R)+H(η)H(Y)\approx \ell H(R)+H(\eta), where HH denotes the high-pass operator. Local variation can therefore reflect both texture and noise. The learned gate combines this cue with infrared directional structure and the spatial prompt to regulate visible detail (Eqs. 13–14).

  • Infrared compensation. Visible darkness and local infrared flatness jointly gate a Laplacian correction, while positive infrared center–surround contrast enters through the spatial prompt. The correction vanishes in locally constant regions and acts on existing infrared variations (Eqs. 18–19).

PIAFusion uses illumination probabilities in its training loss, and PhyFusion uses estimated emissivity and reflectance to guide feature interaction under physical consistency constraints. PDD-Fusion assigns local cues to the visible-detail gate, the low-frequency modality weights, and the infrared residual gain. For example, in Fig. 5(b), a 5×5\times gain applied to the visible input reduced the mean low-frequency visible weight in the glare region from 0.059 to 0.016.

Q2. Choice of frequency decomposition

We chose a single-scale Gaussian split to provide the two quantities used by the fusion branches. The Gaussian low-pass component gives a smooth local intensity baseline, and subtracting it from the input gives a detail residual. The two components sum to the input and share its pixel grid, allowing the spatial weights and gates to act at corresponding locations (Eq. 2).

Appendix B now discusses alternatives. Decimated wavelets and Laplacian pyramids offer multiscale components, which would require reconstruction or alignment and a rule for assigning the additional bands to the branches. Stationary wavelets and Fourier filtering followed by inverse reconstruction can also yield aligned components. The Gaussian split provides one aligned pair per modality for the current architecture. The sensitivity study varies the Gaussian scale within this design.

Q3. Selection and sensitivity of hyperparameters

The exact values are empirical defaults with specific operating roles. The visible brightness midpoint (0.85) sets the half-response near the high end of normalized brightness, and the infrared contrast midpoint (0.30) sets the half-response for positive local contrast. The Gaussian scale (σ=1.5\sigma=1.5) controls the split between the local intensity baseline and the detail residual with a 7×77\times7 kernel. The infrared residual gain (α=0.50\alpha=0.50) scales the gated correction. Appendix C.3 now lists the defaults, explains their roles, and reports one-parameter-at-a-time comparisons.

In a separate fixed-checkpoint test on 64 LLVIP pairs, the largest relative changes in mean EN and MI across the tested settings were approximately 0.81% and 0.58%, respectively. This test characterizes local sensitivity around the reported defaults.

Q4. Qualitative ablations

Figure 6 shows the baseline, intermediate ablation variants, and full model on two LLVIP scenes. Compared with the baseline, the full model makes the background grille more visible in the first scene and the curved colored boundary clearer in the second, while keeping the riders and pedestrians discernible.

Figure 8 compares four prompt settings using aligned crops: no prompt, spatial prompt only, global prompt only, and both prompts. Comparing the no-prompt and spatial-only settings, the illuminated pavement appears less bright with the spatial prompt in this example.

Figure 9 in Appendix C.4 compares the full model with variants that omit one prior group at a time on LLVIP scenes 190545 and 230443. Removing the exposure priors changes local brightness, while removing the reliability priors makes fine details around the vehicle grilles more pronounced. Removing the infrared priors produces subtler changes in these two scenes.

Dear Reviewers,

The authors have posted their rebuttal responses. I would kindly encourage you to review the authors' responses and continue the discussion by providing follow-up comments where appropriate. In particular, please indicate whether the authors' responses have adequately addressed your concerns and update your assessment if necessary.

Thank you for your time and continued efforts!

Your AC

Thank you again for your careful review. We have provided additional explanations and experimental results to address your main concerns and would appreciate any further feedback you may have. If any concerns remain or any part of our responses needs clarification, please let us know. We would be happy to provide further details.

Reviewer_pC8DOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a physics-guided method for the low-light infrared-visible image fusion, aiming to collaboratively achieving information recovery and integration in harsh environments. Specifically, it uses several image derived priors to guide high-frequency enhancement, low-frequency fusion, and infrared compensation, with spatial and global prompts to further refine the fusion process. The provided experiments show that the proposed method effectively enhance the visibility of the fused image.

Critical strengths
  • The degradation-decoupled formulation provides a interesting perspective on low-light infrared-visible image fusion by considering the different degradation characteristics across modalities and frequency components.

  • The proposed mrthod incorporates the physics-inspired priors into the fusion process, guiding exposure correction, structural preservation, and infrared information compensation.

  • The results demonstrate that the proposed mrthod has good cross-dataset generalization, while the lightweight design and efficient inference further support its practical potential.

Critical weaknesses
  • The proposed method shares several ideas with existing work, such as prior-guided fusion, frequency decomposition, and feature modulation. Its main novelty seems to lie in how different degradation cues are assigned to specific fusion operations. It would be better to clarify the core contributions of the proposed method.

  • The degradation cues may not always reflect the intended image properties. For instance, high-frequency variations can arise from both noise and meaningful textures. Experiments with controlled exposure and noise levels would provide stronger evidence for the proposed reliability modeling.

  • A comparison with a similarly sized conventional fusion network under the same training settings would help determine whether the performance gains can be attributed to the proposed degradation-decoupled design rather than increased model complexity.

  • CDDFuse and PhyFusion are closely related to the proposed approach but are missing from the experimental comparisons. Including these methods would make the evaluation more comprehensive.

Questions for authors

Please see the weaknesses section.

6 comments
AuthorsOpenReview ↗

We explain the fusion design and report controlled perturbation tests and a parameter-matched comparison.

We thank the reviewer for the careful reading and constructive suggestions. Changes in the revised manuscript are marked in blue.

Q1. Core contribution

The contribution is to connect each type of degradation to a specific decision in fusion (Sec. 3.2). For high frequencies, visible local variation and infrared structure jointly determine how much visible detail passes through the gate. For low frequencies, exposure cues set the relative contributions of the two modalities, with a high-brightness cue explicitly reducing the visible weight. Where the visible input is dark and the infrared response is locally flat, a joint mask scales a Laplacian residual to enhance existing infrared variations. These are three distinct decisions about detail, modality weighting, and infrared contrast. CDDFuse learns a correlation-driven decomposition of base and detail features. PhyFusion uses radiative-transfer and Retinex priors with physical consistency constraints.

Q2. Controlled exposure and noise

Section 4.4 and Fig. 5 report the branch responses on one LLVIP night scene. We fixed the infrared input and model weights and varied visible brightness (0.1×0.1\times, 1×1\times, 5×5\times) and Gaussian noise (σ=25/255\sigma = 25/255, 50/25550/255, or 75/25575/255) separately. In Fig. 5(a), noise reduced the mean high-frequency gate on flat pavement from 0.683 to 0.622–0.647, while the gate at an infrared-supported pedestrian edge stayed within 0.542–0.569. At the strongest noise level, the gate also fell from 0.565 to 0.506 on tree bark with weak infrared structure. In Fig. 5(b), the low-frequency visible weight fell from 0.654 to 0.494 in a shadow at 0.1×0.1\times and from 0.059 to 0.016 in a glare region at 5×5\times.

Appendix B.1 further tests the texture–noise ambiguity on 64 LLVIP pairs under local darkening and Gaussian noise (σ=10/255\sigma = 10/255). Replacing only the direct spatial variation input to the gate with its spatial mean increased visible-edge transfer from 0.6436 to 0.7388, with nearly unchanged flat-region RMSE (0.04126 vs. 0.04124). Thus, the gate adapts to the tested perturbations, and the spatial cue can also attenuate genuine visible texture.

Q3. Parameter-matched baseline

We retained the two modality encoders and decoder design, removed the degradation branches and prompt encoder, and widened the backbone. The baseline concatenates the two modality features before decoding and has 925,303 parameters versus 902,795 for PDD-Fusion. Both models were trained from scratch with the same LLVIP split, loss, and optimization settings. At epoch 10 with one training seed, PDD-Fusion improved average VIF from 0.7830 to 0.7970 and led on 28 of 30 held-out pairs. MI was similar (2.9795 vs. 2.9801), while the baseline achieved higher EN (6.8640 vs. 6.7674) and QAB/FQ_{AB/F} (0.5730 vs. 0.5696).

Q4. CDDFuse and PhyFusion

Revised Table 1 includes both methods on LLVIP, M$^3$FD, and MSRS. On LLVIP, PDD-Fusion obtains a VIF of 0.552, compared with 0.485 for CDDFuse and 0.551 for PhyFusion. On M$^3$FD, PDD-Fusion achieves higher MI than both methods (4.548 vs. 2.708 and 2.261), while PhyFusion achieves higher QAB/FQ_{AB/F} (0.556 vs. 0.548).

Taken together, the revision connects the design to measurable branch responses and identifies where the gains hold. In particular, the matched-capacity experiment supports a VIF gain at epoch 10, and the controlled tests reveal a trade-off in preserving visible texture.

Dear Reviewers,

The authors have posted their rebuttal responses. I would kindly encourage you to review the authors' responses and continue the discussion by providing follow-up comments where appropriate. In particular, please indicate whether the authors' responses have adequately addressed your concerns and update your assessment if necessary.

Thank you for your time and continued efforts!

Your AC

Thank you again for your careful review. We have provided additional explanations and experimental results to address your main concerns and would appreciate any further feedback you may have. If any concerns remain or any part of our responses needs clarification, please let us know. We would be happy to provide further details.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.