acceptodds
Withdrawn from ICLR 2027

Beyond Sharp Coverage: Focus-Aware Content Routing for Generalized Multi-Focus Image Fusion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multi-focus image fusion (MFIF) combines differently focused images into an all-in-focus result. Conventional methods rely on a sharp-coverage assumption, where each scene region is sufficiently observed in at least one input. In practice, however, limited focal-plane sampling and defocus spread can leave regions blurred in every input, while small viewpoint differences further complicate transferring observed content. We formulate these challenging cases as Generalized MFIF, where source selection alone is insufficient. We then introduce two benchmarks (CommonBlurGeometry and RealScene-68) and a region-wise evaluation protocol. The proposed Focus-Aware Content Routing framework uses continuous focus estimates to route each region according to the availability of reliable sharp observations: it preserves what is reliably observed and restores only where observations are insufficient. A focus-aware router determines the spatial contribution of the source observations and a restored candidate predicted by a one-step pixel-space restorer, while a contextual refiner corrects local inconsistencies in the composite. Our method achieves 32.64 dB on CommonBlurGeometry, exceeding the strongest external baseline by 3.86 dB, leads regional evaluation on RealScene-68, and reaches 42.93 dB on RealMFF. Inference takes 0.190 s at \(625\times433\) and 2.869 s at \(4104\times2736\) per image.

Primary area:
generative models
Keywords:
Multi-focus image fusion, focus-aware routing, selective restoration, image fusion, defocus blur

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_Tv6BOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the challenge in Multi-Focus Image Fusion (MFIF) where blurred regions may exist in both input images. To tackle this, the authors propose Focus Aware Content Routing, which uses a focus map to distinguish between regions that should be preserved from the input images and those that require restoration. The final image is then generated by combining a pre-trained restoration model with contextual refinement. The authors introduce a new benchmark and a region-specific evaluation method, reporting performance that surpasses existing approaches.

Critical strengths
  • The paper addresses the limitations of the complementary focus coverage assumption, a premise common to many traditional MFIF methods, by specifically evaluating restoration performance in scenarios where blurred regions appear in both input images.

  • The overall reconstruction performance is high, and the paper includes ablation studies on key components, as well as visualizations of the routing masks.

Critical weaknesses
  • There is a lack of quantitative comparison with GMFF, the most relevant existing method. GMFF also deals with incomplete focal plane coverage and generative restoration, a fact acknowledged in the Related Work section of this paper. However, the authors do not compare their method against GMFF in their experiments, nor do they explain why it was excluded. Consequently, it is difficult to determine whether the proposed method truly outperforms the most similar existing approach.

  • The Focus Aware Router applies a fixed weight allocation based on predicted focus maps, and the restoration and refinement processes are also built upon existing models. In particular, for CommonBlurGeometry, even when source preservation is excluded, the PSNR reaches 32.12 dB, a figure not significantly different from the 32.64 dB achieved by the full model. Consequently, it remains unclear to what extent the proposed routing mechanism itself contributes to the reported performance gains.

  • The independence of the region wise evaluation is insufficient. The restoration region is defined based on the routing weights predicted by the proposed method, rather than on GT data regarding the actual blur state. While applying the same regions across all baselines is fair, it has not been verified whether these regions actually correspond to areas where both input images are blurred. Given that CommonBlurGeometry includes GT focus information, an evaluation using independently defined regions is feasible.

Questions for authors
  • GMFF is a prior study directly relevant to this work, as it addresses incomplete focal plane coverage. Please explain why this method was excluded from the quantitative comparison and, if possible, provide comparison results using the same two input images and evaluation metrics.

  • When the focus scores for both inputs are low, the weight assigned to the second image also becomes small. As a result, the blended image used by the restoration model relies primarily on the first input image. However, even when both input images are blurred, each may still contain distinct, useful information. Please clarify whether the current input configuration effectively leverages information from both images, compared to an approach that conditions the two original images independently.

  • Please provide a quantitative analysis of how well the predicted routing masks align with the actual sharp regions and jointly blurred regions, using the ground truth focus information from CommonBlurGeometry. Additionally, please verify whether the results presented in Table 2 hold when using ground truth region masks.

  • In Table 4, while the contributions of restoration and contextual refinement are significant, the further improvement gained from source preservation within CommonBlurGeometry appears relatively minor. Please conduct a comparison in which only the routing method is varied, while keeping the pretrained restorer and refiner constant, to more clearly demonstrate the effectiveness of the proposed Focus Aware Content Routing approach itself.

Reviewer_8r7TOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper considers multi-focus image fusion when some regions are blurred in every input and the images have small spatial displacements. A focus estimator controls the contributions of source pixels and a one-step restoration output, followed by a refinement network. The authors introduce CommonBlurGeometry and RealScene-68 and evaluate both whole-image quality and regional preservation and restoration. Experiments cover conventional complementary-focus data, scenes with shared blur, and component ablations.

Critical strengths

1.The paper addresses a practical limitation of source selection: choosing the sharper input cannot recover detail that is blurred in both images. Using the availability of sharp observations to combine source content with restored content is a sensible design. Evaluating preservation and restoration separately also helps explain what the method achieves. 2.The gains in Table 1 are sizeable, with 32.64 dB on CommonBlurGeometry and 42.93 dB on RealMFF. The ablations in Table 4 are useful: restoration contributes most on the shared-blur dataset, whereas preservation and refinement have larger effects on RealMFF. This supports the proposed combination of components.

Critical weaknesses

1.Some experimental details are needed to interpret the comparisons. Appendices A and C (pp. 14–15) give the training and validation set sizes, the scene-disjoint RealMFF split, and some training settings. I could not tell whether Tables 1–2 use these validation sets or whether the same data were used to select checkpoints or hyperparameters. The baseline training procedures also need clarification. For example, Table 6 (p. 18) lists “Flux2 LoRA,” but its fine-tuning data and its relationship to the FLUX2.0 results are not specified. Please describe data usage and model selection for each method so that the effects of different training and adaptation procedures can be assessed. 2.The regional evaluation depends on the method’s own focus estimates. Section 5.2 and Appendix B (pp. 6 and 14) define source-supported and restoration regions from the predicted routing weights. Using the same regions for every method is appropriate. My concern is whether those regions actually correspond to available sharp observations and blur shared by both inputs. This matters because Table 2 is used to support separate claims about preservation and restoration. The synthetic focus supervision in CommonBlurGeometry offers a direct way to check this: define the evaluation regions from those labels and repeat the regional measurements.

Questions for authors

1.Which data are used for the final results in Tables 1–2, and were they also used for checkpoint or hyperparameter selection? Please specify the baseline training or fine-tuning data, particularly for “Flux2 LoRA.” 2.With the models and outputs unchanged, do the preservation gains, restoration gains, and G-RSR improvements in Table 2 hold when the evaluation regions are defined from synthetic focus supervision?

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

This paper studies multi-focus image fusion when some regions are blurred in both inputs and the inputs may have small viewpoint differences. It introduces CommonBlurGeometry and RealScene-68, together with a region-wise evaluation protocol. The proposed method uses focus estimates to combine source observations with a restored candidate, followed by a contextual refiner. Experiments report improvements on the proposed datasets and RealMFF.

Critical strengths

None

Critical weaknesses
  1. The experimental evaluation is insufficient, which is a major weakness of this paper. Only six methods are compared, with only one from 2025 and one from 2026. More recent MFIF methods should be included to establish the claimed advantage. The qualitative comparisons are also limited, and Figure 4 uses DSIFT, published in 2015, as the main comparison on CommonBlurGeometry. In addition, the benchmark details are insufficient. The synthetic data depend on a depth model, but how the benchmark quality is ensured is unclear. For RealScene-68, the capture procedure for the 32 newly collected pairs and their differences from existing datasets are not sufficiently explained. These issues substantially weaken the experimental support for the paper.

  2. The summary and criticism of existing methods are not sufficiently justified, making the motivation unclear. The paper generally describes existing MFIF methods as identifying and combining sharper observations from the inputs. However, this does not necessarily apply to current deep learning methods, especially end-to-end methods that reconstruct the fused image. It is also unclear whether “Conventional MFIF” in Figure 1 includes these methods. Without accurately characterizing what existing methods can and cannot do, the paper does not clearly establish the research gap it aims to address.

Questions for authors
  1. What does “small viewpoint differences” mean in the abstract? Why do such differences arise, and why is it necessary to include them in this task? Please also specify which metric the reported 32.64 dB refers to.

  2. The abstract introduces two benchmarks, while the contributions describe one benchmark. Please clarify this inconsistency.

  3. The synthetic benchmark depends on a depth model. How do the authors ensure the quality of the estimated depth and the resulting benchmark?

  4. What exactly does “The target image is not used for this pairwise alignment” mean in Section 5.1? How are the inputs and the reference image aligned for evaluation?

  5. The compared methods are limited and mostly old, with only one method from 2025 and one from 2026. Can the authors include more recent MFIF methods to substantiate the claimed advantage?

  6. Does “Conventional MFIF” in Figure 1 include end-to-end deep learning methods? What supports the general claim that these methods identify and combine sharper observations from the inputs? This description does not necessarily apply to methods that learn to reconstruct the fused image.

  7. How were the 32 new image pairs in RealScene-68 captured? Please provide the capture settings and procedure, and explain how these pairs differ from existing datasets.

  8. Why does Figure 4 use DSIFT, published in 2015, as the main qualitative comparison on CommonBlurGeometry? Please include direct comparisons with more recent methods on the same examples.

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

Related papers

Discussion (0)

Sign in to comment.