acceptodds
Under review as a conference paper at ICLR 2027

GaIF: Content-Adaptive Patch–Channel Gating for Image Fusion

Abstract

Infrared–visible image fusion requires spatially adaptive modality integration, as various regions demand different balances between thermal saliency and visible texture. However, existing methods often operate at an overly coarse fusion granularity: CNN-based autoencoders merge modality streams before fine-grained routing, while scalar-gated transformers apply a single fusion coefficient uniformly across tokens. This causes a mismatch with the inherently nature of fusion, where modality contributions should vary across patches and feature channels. To address this limitation, we propose GaIF with Patch-Conditioned Modality Routing (PCMR). Bidirectional cross-attention first introduces cross-modal context into into each token, after which a lightweight MLP predicts a gate to enable patch–channel-specific routing at each layer. A confidence-weighted compositor further handles uncertain regions, while an intra-sequence depthwise mixer improves local spatial coherence. Trained only on LLVIP and synthetic multi-focus pairs, GaIF generalizes across IVIF, MEIF, remote sensing, and NIR–VIS fusion tasks, achieving better results than GIFNet except for SCD on Lytro, MFI-WHU, and Harvard datasets. Ablation studies validate each component independently. These results demonstrate the effectiveness and generalizability of fine-grained patch–channel modality routing.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Image Fusion

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

This paper proposes GaIF, a content-adaptive image fusion framework designed to address the limitations of coarse-grained modality integration. The core component, Patch-Conditioned Modality Routing (PCMR), combines bidirectional cross-attention with token-wise and channel-wise gating to dynamically integrate complementary features. Additionally, a Confidence-Weighted Compositor (CWC) and an Intra-Sequence Mixer (TSM) are introduced to improve feature integration and spatial coherence. Experiments across multiple image fusion tasks demonstrate improvements over several existing methods, with ablation studies supporting the effectiveness of the proposed components.

Critical strengths

The paper evaluates the proposed method across multiple fusion tasks and datasets, demonstrating promising cross-task generalization.

The quantitative results and component ablations provide empirical evidence of the effectiveness of the proposed architecture. The inclusion of downstream detection experiments further supports its practical relevance.

Critical weaknesses
  1. Limited technical novelty and insufficient differentiation from existing methods. The proposed framework appears to be an incremental combination of established techniques rather than a fundamentally new fusion paradigm.The primary contribution of PCMR is to extend the gating operation to the patch–channel level through an MLP. However, the paper does not sufficiently explain why this particular combination constitutes a substantial technical advancement over existing adaptive fusion mechanisms.

  2. Insufficient motivation and limited mechanistic justification. The motivation for replacing scalar gating with patch–channel routing is primarily conceptual. The paper lacks convincing failure-case analyses or targeted visual evidence demonstrating the specific limitations of existing gating approaches. Although Figure 5 illustrates different gating granularities, it does not sufficiently establish why fine-grained routing is necessary or how the learned channel-specific gates correspond to meaningful cross-modal information selection.

Questions for authors

What is the fundamental technical distinction between PCMR and existing cross-attention-based adaptive fusion methods, particularly AITFuse and CrossFuse, beyond increasing the granularity of the gating weights?

Can the authors provide representative failure cases of conventional gating methods, together with learned routing maps, to demonstrate why patch–channel-specific fusion is necessary?

Since bidirectional cross-attention accounts for the largest improvement in Table 4, can the authors further isolate the contribution of PCMR using parameter-matched comparisons with alternative adaptive gating mechanisms?

7 comments

W1

We respectfully disagree that PCMR is merely a redundant combination of existing modules. Bidirectional cross-attention exchanges information between the two modality streams; PCMR then uses those cross-attended representations to make an explicit modality-mixing decision for every patch and feature channel. The full factorial ablation in Table 13 measures PCMR’s contribution while holding the other components fixed:

BiCAPCMRTSMCWCEI ↑SCD ↑AG ↑
OffOffOffOff52.141.425.61
OffOnOffOff54.601.475.95
OnOffOffOff58.711.516.10
OnOnOffOff61.831.576.44

With BiCA, TSM, and CWC held fixed, adding PCMR improves EI/SCD/AG by 2.46/0.05/0.34 without BiCA, and by 3.12/0.06/0.34 with BiCA. The result with BiCA is especially relevant to the concern that PCMR merely duplicates cross-modal interaction: PCMR adds a measurable gain after BiCA is already present.

The gate-resolution ablation in Appendix C, Table 5 provides complementary evidence. On LLVIP, PCMR improves all four reported metrics over both scalar and spatial-only gating:

Gate designEI ↑VIF ↑SCD ↑AG ↑
Scalar per-layer gate60.200.711.576.52
Spatial-only gate62.070.751.606.73
PCMR, patch-channel gate64.310.801.636.88

Compared with spatial-only gating, PCMR gains 2.24 EI, 0.05 VIF, 0.03 SCD, and 0.15 AG. These results support PCMR as a distinct, empirically tested routing mechanism. We will make the Table 13 component contrasts and Table 5 gate-resolution results more prominent in the rebuttal and revised manuscript. Since the gate variants are not strictly parameter-matched, we will also add a parameter-matched alternative-gating control and scope our claims to the evidence.

W2

We agree that the manuscript should make the routing behavior easier to inspect. The submission already provides quantitative evidence: Figure 3 compares qualitative fusion outputs, Appendix C, Table 5 evaluates scalar, spatial-only, and patch-channel gates, and Table 13 reports the full component ablation. For example, the LLVIP gate-resolution results are:

Gate designEI ↑VIF ↑SCD ↑AG ↑
Scalar gate60.200.711.576.52
Spatial-only gate62.070.751.606.73
PCMR64.310.801.636.88

These results test the effect of routing granularity, but we agree they do not replace visual examples showing where routing differs in an individual scene. Appendix G currently discusses how to interpret the channel-averaged PCMR gate, but does not display a sample-specific gate map.

In the revision, we will add a compact set of actual inference visualizations to Appendix G. Each example will show the infrared and visible inputs, the scalar-gating and GaIF fused outputs, and the PCMR gate and compositor maps. We will select cases where the outputs reveal a clear difference in preserved thermal targets or visible details, and include a crop to make that difference easier to inspect. The gate map will be identified as the channel-averaged PCMR output, and the compositor map will be described as a learned mixing coefficient, not calibrated uncertainty. This will complement the existing ablations with direct visual evidence, without treating selected examples as a substitute for quantitative evaluation.

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

Q1

All three methods adapt fusion to the input. Their distinction lies in how cross-modal interaction determines the mixture of the two streams.

MethodRole of cross-modal interaction in fusion
CrossFuse [1]Uses cross-attention weights for content-aware gating.
AITFuse [2]Uses an adaptive interactive Transformer to model token- and channel-level interactions.
PCMRRetains both streams through bidirectional cross-attention (BiCA). A separate MLP reads the concatenated infrared and visible tokens at each patch and predicts one coefficient per channel. That coefficient weights the infrared-stream representation and its complement weights the visible-stream representation.

Thus, PCMR's distinction extends beyond gate resolution. Its gate acts after context exchange and explicitly sets the relative contributions of the two contextualized streams. We will clarify this operation in the related-work and method sections. Attention weights determine which tokens contribute to a token representation; PCMR's coefficients determine the mixture of the resulting two streams at each patch-channel coordinate. These operations are sequential but coupled, since the gate depends on cross-attended features.

Appendix D.3, Table 9 compares the complete models under the same LLVIP/TNO protocol.

MethodDatasetEI ↑VIF ↑SCD ↑AG ↑
CrossFuseLLVIP59.840.6921.4876.17
AITFuseLLVIP61.880.7481.5586.47
GaIFLLVIP64.310.8001.6306.88
CrossFuseTNO54.270.9961.6385.06
AITFuseTNO56.271.0131.6635.34
GaIFTNO58.621.0201.6805.55

GaIF scores higher on all four metrics on both datasets. This comparison measures whole-model performance. Within GaIF, Table 4 tests the incremental benefit of adding PCMR after BiCA (Q3 below), and Appendix C, Table 5 tests routing granularity (Q2 below).

Q2

Appendix C, Table 5 compares three gate designs on LLVIP. Even when the gate already varies across patches, allowing the mixture to vary across channels within each patch improves all four reported metrics.

Gate designEI ↑VIF ↑SCD ↑AG ↑
Scalar gate60.200.711.576.52
Spatial-only gate62.070.751.606.73
PCMR64.310.801.636.88

Figure 3 shows fusion outputs, and Appendix G explains channel averaging. To show where the gate designs differ on an individual image, we will add an existing evaluation pair to Appendix G. We will show both inputs, the scalar-gating and GaIF outputs, and matching crops of the region where their outputs differ. Alongside them, we will show the channel-averaged PCMR gate and compositor map from the trained model's forward pass. The caption will specify the channel averaging and identify the compositor values as learned mixing coefficients without uncertainty calibration.

Q3

BiCA gives the largest single EI gain in the incremental ablation (52.14 to 58.71). Appendix F, Table 13 compares PCMR on and off with the other components fixed. The rows below show both BiCA settings with TSM and CWC disabled.

BiCAPCMRTSMCWCEI ↑SCD ↑AG ↑
OffOffOffOff52.141.425.61
OffOnOffOff54.601.475.95
OnOffOffOff58.711.516.10
OnOnOffOff61.831.576.44

With BiCA present, adding PCMR raises EI/SCD/AG from 58.71/1.51/6.10 to 61.83/1.57/6.44, gains of 3.12/0.06/0.34. Thus, PCMR improves the evaluated model even after the two streams have exchanged information. Table 5 also shows gains over both coarser gates. The comparison with spatial-only gating is particularly relevant: it tests the benefit of allowing channels within the same patch to use different input-dependent modality mixtures, although gate capacity is not matched. These comparisons control which modules are present, but gate parameter counts differ. We will state explicitly that the measured gain applies to the PCMR design as implemented. Isolating how much comes from routing granularity alone would require a parameter-matched control.

References

[1] H. Li and X.-J. Wu, “CrossFuse: A Novel Cross Attention Mechanism Based Infrared and Visible Image Fusion Approach,” Information Fusion, 2024.

[2] Z. Wang, F. Yang, J. Sun, J. Xu, F. Yang, and X. Yan, “AITFuse: Infrared and Visible Image Fusion via Adaptive Interactive Transformer Learning,” Knowledge-Based Systems, vol. 299, Art. no. 111949, 2024.

Reviewer_Hq3jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GaIF, which combines bidirectional cross-attention with patch–channel gating for content-adaptive fusion, and evaluates the architecture across multiple image fusion tasks.

Critical strengths

The motivation is clear, and PCMR provides an intuitive, fine-grained fusion mechanism that addresses varying modality requirements across spatial regions and feature channels. Trained only on LLVIP and synthetic multi-focus data derived from it, the proposed method shows evidence of transferability across multiple datasets and fusion tasks. The ablation studies are reasonably comprehensive and provide empirical support for the effectiveness of the main design choices.

Critical weaknesses

The authors identify patch–channel gating after bidirectional cross-attention as the key distinction from AITFuse. However, the current experiments do not adequately disentangle the benefits of this design from those of enhanced feature interaction and greater model capacity, leaving its independent contribution insufficiently established. In Eq. (3), the FFN is applied separately to the two modality streams, whereas Figure 2 appears to apply it to the fused representation. These descriptions imply different computational paths. The main text describes patch embedding as a learnable P×PP\times P convolution, yet Table 7 reports the same parameter count of 14.8M for different values of PP. Under a standard convolutional implementation with the rest of the architecture unchanged, should the parameter count not vary with the kernel size? Minimizing the weighted IVIF MSE term in isolation yields a weighted average of the source images with spatially uniform weights as the optimal output. This is in tension with the stated objective of region-specific modality preservation. Although the full objective also includes SSIM and auxiliary supervision, how these components jointly support local modality selection remains insufficiently analyzed. TSM mixes tokens using a one-dimensional convolution. If the tokens are flattened in row-major order, this operation may connect spatially distant patches across row boundaries and does not necessarily reflect two-dimensional local neighborhoods. Its purported role in enhancing two-dimensional spatial continuity is therefore insufficiently substantiated.

Questions for authors

Please see Critical Weaknesses.

4 comments

W1

The component ablation compares BiCA alone with BiCA + PCMR, keeping cross-attention fixed. Adding PCMR improves LLVIP EI/SCD/AG by 3.12/0.06/0.34:

ConfigurationEI ↑SCD ↑AG ↑Change from previous row
Symmetric-residual baseline52.141.425.61—
+ Bidirectional cross-attention (BiCA)58.711.516.10+6.57 / +0.09 / +0.49
+ PCMR61.831.576.44+3.12 / +0.06 / +0.34
+ Intra-sequence mixer (TSM)63.091.616.68+1.26 / +0.04 / +0.24
+ Confidence-weighted compositor (CWC)64.311.636.88+1.22 / +0.02 / +0.20

PCMR thus adds a gain after cross-modal interaction. Since variants are not parameter-matched, routing granularity and gate capacity are not fully separated. We will state this limitation and add a parameter-matched alternative-gating control, reporting parameters and results.

W2

Equation (3) and Fig. 2 describe different FFN paths. We will check the implementation and align the equation, text, and figure, specifying whether the FFN updates the modality streams or the fused token. The reported experiments use the implemented model.

W3

Appendix D.1, Table 7 reports patch-size sensitivity on 128 × 128 inputs under the same training and evaluation protocol:

Patch sizeTokensParams (M)FLOPs (G)Memory (GB)Latency (ms)EI ↑VIF ↑SCD ↑AG ↑
P = 4102414.812.08.022.064.000.791.626.84
P = 825614.84.83.29.564.310.801.636.88
P = 166414.83.21.87.062.800.771.606.70

P = 8 gives the best measured accuracy-computation trade-off. P = 4 quadruples tokens and increases memory/latency; P = 16 reduces computation but lowers detail-sensitive scores.

The identical 14.8M counts need verification because the stated learned P × P convolution changes with P. We will recount parameters from the implementation, including patch embedding and all projections, then correct Table 7 or clarify the counting convention. Identical counts will be retained only if verified.

W4

With global source weights, the isolated weighted-MSE optimum is a spatially uniform weighted blend. This term alone does not directly supervise patch-level PCMR gates.

Appendix D.2, Table 8 evaluates the training objectives and activity priors:

SettingActivity priorEI ↑VIF ↑SCD ↑AG ↑
Full GaIFDenseNet-12164.310.8001.6306.88
VGG activityVGG-1664.060.7961.6246.83
Uniform source weightsNone63.680.7891.6196.79
Without reconstructionDenseNet-12163.340.7841.6136.70
Without token alignmentDenseNet-12163.860.7941.6266.82
Without confidence regularizationDenseNet-12163.980.7961.6236.85
Without auxiliary regularizersDenseNet-12163.570.7881.6176.76
IVIF-only trainingDenseNet-12160.910.7611.5586.53

The full training scheme includes MSE, SSIM, reconstruction, token alignment, confidence regularization, and alternating MFIF supervision. The table shows performance drops from removing individual terms; removing MFIF causes a larger EI/AG drop, from 64.31/6.88 to 60.91/6.53. These results support the auxiliary objectives and MFIF supervision, but do not establish direct local gate labels. Routing benefits are tested in Tables 4 and 5. We will distinguish global loss weights from local gate supervision and clarify that MSE alone does not explain spatially varying routing.

W5

TSM uses a depthwise 1-D convolution with kernel size 3 on flattened tokens. Sequence neighbors need not be 2-D neighbors: a row's last token can become adjacent to the next row's first token.

The component ablation measures TSM's incremental contribution:

ConfigurationEI ↑SCD ↑AG ↑Change from previous row
+ Bidirectional cross-attention58.711.516.10—
+ PCMR61.831.576.44+3.12 / +0.06 / +0.34
+ TSM63.091.616.68+1.26 / +0.04 / +0.24
+ CWC64.311.636.88+1.22 / +0.02 / +0.20

TSM improves results after PCMR, but this does not prove 2-D continuity. We will replace claims of enforcing/restoring 2-D spatial continuity with “encourages local coherence between neighboring positions in the flattened token sequence.”

Where possible, we will compare a 2-D depthwise-convolution control using the same input size, training schedule, and evaluation protocol to distinguish general local-mixing benefits from those of the sequence-based implementation.

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 for the response. I view the use of patch–channel gating after bidirectional cross-attention as a useful but incremental extension of prior work such as AITFuse. The ablations show that adding PCMR improves fusion performance, but without a parameter-matched gating comparison, it remains unclear how much of the gain comes from the finer routing granularity. I therefore maintain my original rating.

Thank you for recognizing the practical value of PCMR. We agree that quantifying how much of the gain comes from routing granularity alone requires a parameter-matched comparison. The current results support a narrower claim about the design, and we will state that claim more precisely in the manuscript.

AITFuse [1] uses token-wise and channel-wise attention to model feature interactions. Selective Kernel Networks [2] also established a precedent for channel-wise selection among convolutional branches. Our contribution concerns which features the gate reads and which decision its output controls. After bidirectional cross-attention, GaIF retains two contextualized feature vectors at each patch. A shared MLP reads both vectors and predicts a coefficient gp,cg_{p,c}. This coefficient and its complement 1−gp,c1-g_{p,c} directly determine the relative contributions of the two streams at patch pp, channel cc.

Consider a patch containing both a pedestrian’s thermal outline and visible clothing texture. If different channels encode these cues, PCMR can assign a larger infrared-stream weight to one channel and a smaller weight to another. A spatial-only gate must apply the same ratio to both. Increasing the capacity of a predictor that still outputs one scalar per patch cannot remove this constraint at the mixing step. Conversely, tying PCMR’s coefficients across channels recovers the spatial-only mixing rule. This identifies precisely what finer routing makes possible, independently of how much it improves the measured scores.

The closest empirical comparison is Appendix C, Table 5, because the spatial-only gate already adapts its weights across patches. On LLVIP, the results are:

Gate designEI ↑VIF ↑SCD ↑AG ↑
Scalar60.200.711.576.52
Spatial-only62.070.751.606.73
PCMR64.310.801.636.88
PCMR gain over spatial-only+2.24+0.05+0.03+0.15

Appendix F, Table 13 addresses a related question: does adding PCMR still help once bidirectional cross-attention is already present? With cross-attention enabled in every comparison, adding PCMR raises EI by 3.12, 3.10, 3.54, and 4.16 with neither TSM nor CWC enabled, CWC only, TSM only, and both enabled, respectively. Thus, the gain appears in all four configurations, including the model that already contains both other modules (60.15 to 64.31 EI).

We will revise Sec. 2.3 to make the routing operation explicit and the ablation discussion to reflect the evidence precisely. The evaluated PCMR design improves on spatial-only gating across all four LLVIP metrics. The fraction of that improvement attributable to granularity alone remains an open question.

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

This paper proposes a dynamic infrared–visible image fusion framework, termed GaIF. By introducing Patch-Conditioned Modality Routing (PCMR), GaIF adaptively regulates the contributions of different modalities across image regions and feature channels, thereby achieving fine-grained image fusion.

Critical strengths
  1. The paper presents fusion results that are visually distinct from those of existing image fusion methods.
Critical weaknesses
  1. The novelty of PCMR compared with existing joint spatial-channel attention and dynamic gating mechanisms is insufficiently demonstrated.
  2. The experimental section lacks visualizations of the learned gating maps.
  3. The manuscript lacks comparisons with recent representative methods published in 2026.
Questions for authors

1.What are the key differences between the proposed method and existing image fusion methods based on channel attention and spatial attention? Such methods can also achieve fine-grained content-adaptive fusion at both the channel and spatial levels. 2. It is recommended that the authors further analyze the differences and advantages of PCMR compared with existing joint spatial-channel attention and dynamic gating mechanisms. 3. It is recommended that the authors provide visualizations of the actual learned gating maps to support the claimed fine-grained modulation behavior. 4. It is recommended that the authors include comparisons with recent representative methods published in 2026. 5. Are the feature maps and confidence maps shown in Figs. 1 and 2 obtained from actual model outputs or are they illustrative visualizations?

5 comments

W1

We respectfully disagree with the assessment that PCMR is insufficiently differentiated from existing joint spatial-channel attention and dynamic gating mechanisms. PCMR introduces an explicit cross-modal-context-conditioned routing operation at patch-channel resolution. Bidirectional cross-attention first enriches both modality streams with information from the other modality. A dedicated MLP then jointly reads the cross-attended token pair and predicts Gn,dG_{n,d}, which directly controls the complementary mixture

Fn,d=Gn,dT^A,n,d+(1−Gn,d)T^B,n,d.F_{n,d}=G_{n,d}\hat{T}_{A,n,d}+(1-G_{n,d})\hat{T}_{B,n,d}.

This design separates cross-modal information exchange from the subsequent modality-mixture decision: each patch can select a different mixture for each feature channel, conditioned on both contextualized streams.

The contribution of this routing operation is supported by two complementary ablations. In Table 4, adding PCMR after bidirectional cross-attention increases EI/AG from 58.71/6.10 to 61.83/6.44. In Table 5, patch-channel routing outperforms spatial-only gating on all four LLVIP metrics: EI/VIF/SCD/AG increase from 62.07/0.75/1.60/6.73 to 64.31/0.80/1.63/6.88. These results support both the additional benefit of explicit routing after cross-attention and the value of channel-specific modality mixing within each patch.

W2

The submission already includes a gate-resolution visualization in Appendix E, Fig. 5 (p. 16), comparing scalar, spatial-only, and patch-channel routing alongside fused outputs and enlarged local crops. Appendix G (p. 17) explains how the channel mean of the N×DN\times D PCMR gate can be reshaped into a spatial routing summary. The corresponding quantitative comparison is provided in Appendix C, Table 5 (p. 13).

These materials provide visual and quantitative evidence concerning routing granularity. We recognize that the request specifically concerns actual learned gate maps, whose provenance is not sufficiently documented in the current figure. We will clarify the visualization procedure and figure caption, explicitly identify schematic elements and inference-derived maps, and make these existing appendix references more prominent in the main text.

W3

Among the 2026 studies we examined, publicly reported results directly comparable under our LLVIP/TNO evaluation protocol are limited. For example, Customized Fusion [1] evaluates fusion quality on M3FD, FMB, and VT5000, while Degradation-Robust Fusion [2] evaluates fusion under noise, blur, and composite degradation. Their published scores therefore cannot be directly inserted into our LLVIP/TNO benchmark as same-protocol comparisons.

Our submitted manuscript already includes quantitative comparisons with CrossFuse, ATFuse, AITFuse, YDTR, PPT Fusion, and GIFNet on LLVIP and TNO using EI/VIF/SCD/AG (Appendix D.3, Table 9). These comparisons cover several adaptive and Transformer-based fusion approaches directly relevant to PCMR. The two aforementioned 2026 papers are also cited in Sec. 2.3.

We will expand Sec. 2.3 to explicitly compare these recent methods with GaIF in terms of fusion mechanism, supervision, and evaluation setting. We will also prioritize adding a representative 2026 baseline with released code and checkpoints that permit evaluation under the same test split, preprocessing, and metric implementation. Results quoted from the original papers will be clearly identified as published results rather than our own re-evaluations.

References

[1] Zengyi Yang, Yu Liu, Juan Cheng, Zhiqin Zhu, Yafei Zhang, and Huafeng Li. Customized Fusion: A Closed-Loop Dynamic Network for Adaptive Multi-Task-Aware Infrared-Visible Image Fusion. CVPR, 2026.

[2] Yu Shi, Yu Liu, Zhong-Cheng Wu, Juan Cheng, Huafeng Li, and Xun Chen. Degradation-Robust Fusion: An Efficient Degradation-Aware Diffusion Framework for Multimodal Image Fusion in Arbitrary Degradation Scenarios. CVPR, 2026.

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

Reviewer_MnwVOpenReview ↗

Response to the Rebuttal

Thank you for the detailed response. The authors have addressed some of my concerns. I recommend that they clarify the origin of the actual learned gating maps and include a more substantive discussion of representative methods published in 2026 in the manuscript.

However, I still have the following concerns:

  1. The fundamental distinction between the proposed method and existing joint spatial–channel attention mechanisms remains unclear.
  2. The formulation presented in the response to W1 suggests that the gating mechanism inherently involves both cross-modal information interaction and modality fusion, rather than explicitly separating these two processes as claimed by the authors.
  3. There appear to be garbled mathematical symbols in the bottom-right corner of Fig. 2, which should be corrected.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.