acceptodds
Under review as a conference paper at ICLR 2027

Decision-Centric Fusion: A Unified Decision Variable for Modality Alignment and Task Adaptation

Abstract

Infrared and visible image fusion is fundamentally limited by modality imbalance, where visible textures degrade under adverse conditions while infrared remains stable. Existing methods address this by optimizing encoding, fusion, and refinement as independent modules, yet they lack a shared degradation consensus, causing information asymmetry that undermines fusion quality. We propose Decision-Centric Fusion, which introduces a unified decision variable as the global decision signal, enabling all modules to optimize within the same degradation-aware space. The framework comprises three native components. The System Decision Module generates the unified variable to guide the full pipeline. Building on this, the Zip-Alignment Module aligns the two modalities through four interlocking gears: encoder attention alignment with infrared high-frequency injection, adaptive rank decomposition, soft query replacement, and decision-guided fusion weighting. The Boundary-Aware Refiner then identifies a safe operating region, ensuring that post-processing improves perceptual quality without degrading task performance. To facilitate deployment, knowledge distillation compresses the pipeline to 0.25M parameters with negligible loss. Experiments on four benchmarks validate the effectiveness, efficiency, and boundary behavior of the proposed framework.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal, Infrared and Visible Image Fusion, Decision theory

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ppXaOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Aiming at the problem that the spatial reliability of visible infrared image fusion changes in the image, this paper proposes a decision-making centered fusion method DCFusion, which makes the idea of dual-mode alignment not only stay in an isolated stage of image fusion, but also play a role in the whole process. Based on CNN and other basic frameworks, the author uses a variety of visible and infrared datasets for experiments, and proposes System Decision Module, Zip-Alignment Module and Boundary-Aware Refiner, which achieve better experimental results and model distillation without too much performance loss under smaller conditions.

Critical strengths

1. The motivation and proposed solutions of the paper are highly consistent, and they basically solve the problems raised.

2. The experimental design is very complete. The comparison experiments, parameter experiments, detection experiments, and segmentation experiments all have clear descriptions of their settings.

3. The authors propose DCFusion, a lightweight model that ensures both fusion performance and a certain degree of robustness.

4. The experiments involve multiple datasets, and random seed experiments are conducted to demonstrate the transparency of the paper's results.

5. The proposed Zip-Alignment Module achieves sufficient modal information interaction while remaining lightweight.

Critical weaknesses

1. The paper is not clear enough in its expression. Some statements, such as “different modules may operate under different spatial reliability conditions,” generally emphasize that they will select appropriate modules under suitable spatial reliability conditions, but the paper as a whole does not convey this meaning.

2. The figures in the paper are not vector graphics. When presenting qualitative results, such figures are much less convincing.

3. The paper has some formatting issues. In the citation part, the author uses ~\cite{} throughout, but in the author’s context, these citations are not used as the subject, and ~\citep{} should be used for citation. There are also some missing spaces in the paper; some are caused by incorrect citations, and some may be caused by carelessness.

4. The high- and low-frequency decomposition used in Section 3.2 is mainly based on Avgpool. Such high- and low-frequency decomposition is too coarse and may affect the performance.

5. The textual description in Section 3.3 does not correspond to Figure 2, which causes some difficulty in understanding.

6. There are some problems with the explanation of the metrics in Section 4.2. Since metrics such as PSNRPSNR are reference-based image fusion metrics, visible and infrared images need to be provided during calculation. Under the author’s assumption, these data may already have some degradation. In this case, how the metrics calculated can prove that the proposed method has the ability to cope with degradation remains open to question.

7. This paper considers that degradation mainly exists in the visible modality, so when calculating GG, only visible-modality images are calculated. However, the paper also mentions that infrared images have some contrast insufficiency problems, and the expression on this issue needs to be strengthened.

8. In Section 3, this paper has a Degradation perception module, but the entire pipeline faces relatively clean datasets. We can only roughly perceive the role of this module in the ablation study. Apart from that, the capability of this Degradation perception module is difficult to quantify.

9. There is a clear conflict between the paper's main results and Table 11 in the appendix. On the FMB dataset, the main table reports PSNRPSNR for "Ours" as 18.03, SCDSCD as 1.59, and mIoUmIoU as 0.507, whereas Table 11 reports the mean over three random seeds as PSNRPSNR 17.703 ± 0.096, SCDSCD 1.562 ± 0.004, and mIoUmIoU 0.4707 ± 0.0010. If Table 11 summarizes the same FMB experiment, then 18.03 differs from 17.703 by about 3.4 reported standard deviations, 1.59 differs from 1.562 by about 7 reported standard deviations, and the mIoU difference between 0.507 and 0.4707 is even larger.

Questions for authors

1. I do not understand the meaning of the sentence “Different modules may operate under different spatial reliability conditions, even when the network is trained end-to-end.” In particular, I am not sure what “operate” specifically means here, and the authors may need to explain it further. This is because “different modules” generally emphasizes that we would select appropriate modules under suitable spatial reliability conditions, but the paper does not convey this meaning throughout.

2. The high- and low-frequency decomposition used in Section 3.2 mainly relies on AvgPool. Such a high-/low-frequency decomposition is too coarse and may affect the performance. The authors could consider replacing it with wavelet transform or high-/low-pass filters, etc., to demonstrate that the proposed module indeed works in different frequency bands.

3. The figures in the paper are not vector graphics. When presenting qualitative results, perhaps the authors could consider replacing them with vector graphics?

4. The textual description in Section 3.3 does not correspond to Figure 2, which causes some difficulty in understanding. How are IirI_{\text{ir}} and GG fused with IFuseI_{\text{Fuse}}?

5. Given points 6 to 8 in the Critical Weaknesses, would the authors consider conducting tests on some degraded datasets, such as the EMS dataset[1] or the DeMMI-RF dataset[2], to fully demonstrate the effectiveness of the model?

6. In your citation, RealPortrait: Realistic Portrait Animation with Diffusion Transformers appears. Its subject is portrait animation and diffusion Transformers, rather than infrared-visible image fusion.

7. The conflict between the paper's main results and Table 11 in the appendix is confusing. Could the authors please clarify this discrepancy for me?

[1]: Yi X, Xu H, Zhang H, et al. Text-if: Leveraging semantic text guidance for degradation-aware and interactive image fusion[C]//2024 IEEE/CVF conference on computer vision and pattern recognition (CVPR). IEEE, 2024: 27016-27025.

[2]: Sun Y, Li X, Zhu P, et al. Task-gated multi-expert collaboration network for degraded multi-modal image fusion[C]//International Conference on Machine Learning. PMLR, 2025: 57571-57586.

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

This work proposes a decision-centric fusion method that seeks to address the problem of delayed cross-modal information transfer through early bidirectional alignment and persistent spatial decision sharing. The overall idea reflects thoughtful consideration, however, the presentation and implementation of the proposed method may not fully align with the motivation articulated in the paper.

Critical strengths

One strength of this paper is its in-depth understanding of the image fusion problem, which enables the authors to identify new research questions within a well-established field and demonstrates thoughtful consideration of the problem. In addition, the technical roadmap is clearly presented, the writing generally adheres to academic conventions, and the experimental evaluation is relatively comprehensive.

Critical weaknesses

One clear concern is that the final technical approach does not fully align with the motivation of the paper. Although the work aims to address the limitations of existing methods, the proposed design still relies on techniques commonly adopted by these methods and does not appear to depart substantially from the existing design paradigm. As a result, the distinction between the proposed method and existing approaches remains insufficiently clear.

Questions for authors
  1. The paper argues that cross-modal information interaction in existing fusion methods occurs relatively late. However, as shown in Figure 1, feature interaction in the proposed method also takes place after feature extraction, which is similar to the design of many existing approaches. Therefore, it is unclear how the proposed method differs from existing methods in terms of the timing of cross-modal interaction.
  2. Section 2.3 compares the proposed method with existing approaches. However, the preceding sections do not provide a sufficiently detailed description of the proposed method, particularly in the final paragraph of the Introduction. As a result, it is difficult to understand the fundamental differences between the proposed method and existing approaches.
  3. The implementation of the Gear 1 module in Section 3.2 is essentially a cross-modal attention mechanism. This type of interaction has been widely adopted in the image fusion literature, making it difficult to argue that the proposed interaction process differs substantially from those used in existing methods.
  4. The novelty of the Self-Adaptive Rank Decomposition module is unclear. Eqs. (9)–(13) mainly combine average-pooling-based feature decomposition, convolutional branches with different channel bottleneck widths, and gate-based feature mixing. It remains unclear how this design differs substantially from conventional feature decomposition and gated fusion methods.
  5. In Fig. 3, the regions corresponding to the zoomed-in patches are not marked in the original images, and the enlarged input patches appear to be presented incorrectly. Furthermore, since the proposed method does not include any specific designs for downstream tasks, it is unclear why it substantially outperforms the competing methods on these tasks.
Reviewer_kYtiOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Decision-Centric Fusion (DCFusion) for infrared-visible image fusion under spatially varying modality reliability. DCFusion estimates a spatial gating map from the visible image, enhances modality reliability via bidirectional cross-modal alignment, and further reduces the model size using model distillation. The authors report fusion metrics on four public datasets and perform evaluations on downstream semantic tasks.

Critical strengths
  1. The research problem is relatively meaningful. Existing fusion models lack effective strategies for judging modality reliability and corresponding evaluation protocols.

  2. The evaluation framework is comprehensive. The authors consider not only the visual quality of fused images but also conduct assessments from the perspective of downstream semantic tasks, including object detection and semantic segmentation.

Critical weaknesses
  1. The novelty of this paper is limited. Existing studies have extensively explored adaptive techniques for cross-modal perceptual fusion, such as IVIFormer (ICCV 2025), DAFusion (AAAI 2026), and RegionFuse (CVPR 2026). The primary distinction of this work appears to be sharing a single spatial gating map across multiple operations. However, the authors have not demonstrated that such sharing delivers a substantial conceptual or performance advance, and no closely related methods are included for comparison.

  2. The role of the unified decision variable is not sufficiently justified. The paper mainly describes how gradients flow to the shared variable, which does not prove that shared gating necessarily yields more reliable decisions or better solutions than separate gating. Moreover, Gear 1 is not directly modulated by G, which is inconsistent with the claim of "cross-stage shared decision making". The ablation study on separate gating in Table 3 offers partial support, yet controlled comparisons under identical parameter and computation budgets, together with direct quantitative evidence for decision consistency, are still required.

  3. The baseline training protocol and the reported results raise serious reliability concerns. Although Appendix A.2.1 provides the numbers of training, validation and test samples used by the authors, the paper does not clarify whether all ten compared fusion methods are retrained under the identical train-validation-test split, nor does it specify the training samples contained in the adopted public weights. Meanwhile, the mIoU of FMB is 0.507 in Table 2 while it becomes 0.4707 ± 0.0010 in Table 11. Similarly, the PSNR values are 18.03 and 17.703 ± 0.096 in Table 1 and Table 11, respectively. Such substantial variance induced by random seeds raises concerns about the convergence of the models.

  4. The use of PSNR as evidence of fusion fidelity is insufficiently justified. There exists no fused ground-truth image for infrared-visible image fusion, whereas PSNR measures the pixel-level similarity between the output and a specified reference image. It is unreasonable for the authors to adopt PSNR as a comparison metric for fusion, and the rationale behind this choice requires further clarification.

Questions for authors

Please refer to Critical Weaknesses for details.

1 comment

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

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.