acceptodds
Under review as a conference paper at ICLR 2027

Trajectory-aware Differentiable Blur Augmentation for Image Deblurring

Abstract

Image deblurring aims to recover sharp images degraded by motion blur caused by camera shake and object motion. Since collecting blur-sharp pairs covering diverse motion patterns is costly, blur augmentation provides an effective way to expand the training distribution, particularly under limited training data. However, existing methods primarily diversify blur patterns through random perturbations of blur magnitude and orientation, resulting in limited exploration of diverse motion trajectories underlying real-world blur formation. To overcome this limitation, we propose a differentiable trajectory-aware blur augmentation framework that explicitly expands motion-space coverage through trajectory modeling. Specifically, we introduce a Global Trajectory-aware Augmentation (GTA) module that parameterizes camera motion using motion dynamics and trajectory primitives to synthesize diverse motion trajectories. Furthermore, we propose a Local Trajectory-aware Augmentation (LTA) module to model spatially varying local motion, thereby capturing diverse local blur variations and enriching blur diversity. The differentiable formulation enables trajectory augmentation to be optimized through the deblurring objective. Experimental results demonstrate improved blur diversity and deblurring performance across both standard and limited-data training settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Deblurring, Blur Augmentation, Trajectory-aware Blur Augmentation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_A4pVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a Trajectory-Aware Differentiable Blur Augmentation framework for improving image deblurring, particularly under limited training data. The central motivation is that existing blur augmentation methods mainly manipulate motion magnitude and orientation without explicitly modeling the temporal dynamics and geometric structure of motion trajectories. The proposed framework contains two components. First, Global Trajectory-aware Augmentation (GTA) models camera motion using velocity, acceleration, curve and S-curve primitives, and in-plane rotation. Second, Local Trajectory-aware Augmentation (LTA) represents spatially varying residual motion through learned local trajectory components and spatial mixing weights. A differentiable blur renderer synthesizes augmented images, while a learnable Augmentor is optimized using the deblurring objective to generate challenging trajectory perturbations.

Critical strengths

The paper presents a well-motivated and technically reasonable trajectory-aware blur augmentation framework that addresses the limitations of conventional magnitude- and orientation-based augmentation methods. By combining Global Trajectory-aware Augmentation (GTA) and Local Trajectory-aware Augmentation (LTA), the proposed approach effectively models diverse temporal motion dynamics and spatially varying blur patterns. The differentiable augmentation strategy further enables trajectory perturbations to adapt to the deblurring objective, providing a meaningful contribution beyond predefined random augmentation. Experimental results demonstrate consistent improvements across multiple datasets, deblurring architectures, and training-data regimes, with particularly encouraging gains under limited-data settings. The component ablations and qualitative visualizations further support the effectiveness of the proposed framework, while the clear presentation and detailed implementation settings enhance the overall quality of the paper.

Critical weaknesses

Although the proposed method achieves consistent performance improvements across multiple experiments, several aspects could be further strengthened. First, the trajectory diversity analysis relies on the authors' own trained trajectory estimation network, which may introduce evaluation bias and does not sufficiently validate the physical realism of the generated motion trajectories. Second, the additional performance gains brought by Local Trajectory-aware Augmentation (LTA) and the differentiable augmentation strategy are relatively modest, and their individual contributions require further investigation. Moreover, the framework requires pretraining a reblurring network and adopting an alternating optimization strategy, which increases training complexity, yet the associated computational overhead is not quantitatively evaluated. Finally, the comparison with GeoSyn mainly relies on previously published results rather than a fully unified experimental protocol, which somewhat limits the conclusiveness of the reported performance advantages.

Questions for authors
  1. Can the authors provide an independent evaluation to validate the diversity and physical realism of the generated motion trajectories?
  2. Can the authors further clarify the individual contributions of Local Trajectory-aware Augmentation (LTA) and the differentiable augmentation strategy to the overall performance improvements?
  3. Can the authors provide a quantitative analysis of the additional training time, computational cost, and GPU memory overhead introduced by the proposed framework?
  4. Can the authors further clarify the fairness and consistency of the experimental comparison with GeoSyn?
  5. Can the authors further evaluate the generalization ability of the proposed method under complex real-world degradations, such as rolling-shutter effects and sensor noise?
4 comments

We thank the reviewers for their constructive feedback and address their concerns below. Newly added text and figure/table captions in the revised manuscript are temporarily highlighted in blue\textcolor{blue}{\text{blue}}, while table contents remain black for readability. We have also condensed existing text to accommodate additional experiments and welcome further discussion.


[A1] Independent Evaluation of Trajectory Diversity and Blur Realism

We appreciate the reviewer's concern regarding the independent validation of trajectory diversity and generated blur realism. To address this, we conduct an additional trajectory diversity analysis using Motion-ETR [1], an independent estimator that models motion trajectories over the exposure interval.

As shown in Figure 7 and discussed in Section 4.4 of the revised manuscript, the independent analysis reveals diverse trajectory clusters in the original GoPro dataset, consistent with our original findings. Existing augmentation methods remain concentrated in a limited number of clusters, whereas our method achieves broader coverage across different trajectory patterns.

Our primary objective is to improve downstream deblurring performance by adaptively generating diverse blur patterns tailored to the training dataset and target network. To further evaluate the generated blur, we measure recall and coverage for diversity, along with FID and CMMD for distributional similarity. As shown in Table 4 of the revised manuscript (reproduced below), our method outperforms ID-Blau across all four metrics. Although these results do not directly establish physical realism, they suggest that our method generates diverse blur patterns without deviating substantially from the original GoPro blur distribution.

MethodFID↓CMMD↓Recall↑Coverage↑
ID-Blau26.490.0940.9440.978
Ours25.960.0350.9700.984

[A2] Contributions of LTA and Differentiable Augmentation

We appreciate the reviewer's concern regarding the individual contributions of LTA and differentiable augmentation. To further clarify their effects, we expanded our ablation study by including an additional GTA-only setting with random trajectory perturbations.

As shown in Table 3 of the revised manuscript (reproduced below), the expanded ablation enables direct comparisons between random and differentiable augmentation, both with and without LTA, allowing us to assess the individual contributions of local trajectory modeling and feedback-driven optimization. The results show consistent improvements from adding LTA and applying differentiable augmentation across GoPro and HIDE, with the full configuration achieving the best performance across all metrics. Together, these results support the complementary roles of LTA and differentiable augmentation in improving downstream deblurring performance.

GTALTADifferentiableGoProHIDE
MUSIQ↑LPIPS↓PSNR↑DISTS↓MUSIQ↑LPIPS↓PSNR↑DISTS↓
38.490.15528.920.09743.640.19226.710.113
✓42.780.10831.080.08049.940.15528.630.088
✓✓43.200.10431.240.07750.640.15028.690.085
✓✓42.390.10631.190.07750.760.14728.780.083
✓✓✓43.420.09931.370.07451.370.14228.980.081

[A3] Computational Overhead

We appreciate the reviewer's concern regarding the additional computational cost introduced by reblurring network pretraining and alternating optimization. To clarify this overhead, we compare estimated training time and measured peak GPU memory usage for the baseline, ID-Blau, and our method on GoPro using MIMO-UNet+ as the deblurring backbone. The results are reported in Table 8 of the revised manuscript (reproduced below).

For this cost comparison, we follow the original MIMO-UNet+ training protocol for all three methods. We separately report the estimated GPU-hours for reblurring pretraining, deblurring pretraining, and main training. Training costs are estimated from step times measured on a single NVIDIA RTX 3090 with BF16 mixed precision. ID-Blau additionally uses 5,000 epochs of diffusion training and 500 epochs of deblurring pretraining on a generated dataset five times the size of the original training set. The reported total includes both pretraining stages and excludes offline dataset generation for ID-Blau. Our method uses 1,000 epochs of reblurring pretraining, and its main-training time includes online Augmentor optimization. As shown in Table 8, our method achieves a comparable estimated total training time to ID-Blau, although it requires higher peak GPU memory due to batch concatenation.

Aug.Estimated training time (GPU-hours)↓Peak memory (GB)↓
Reblur pretrainingDeblur pretrainingMain trainingTotal
–––66.9666.966.31
ID-Blau13.9155.9066.96136.776.31
Ours9.59–149.70159.2913.03

[A4] Fairness of the Comparison with GeoSyn

We appreciate the reviewer's concern regarding the fairness of our comparison with GeoSyn. As clarified in Appendix A.2, GeoSyn's official code and pretrained blur synthesis model are not publicly available, preventing us from reproducing its results under consistent experimental settings or conducting additional comparisons across different backbones and datasets. We therefore report the RSBlur result from the original GeoSyn paper as a reference, rather than as a direct comparison under identical experimental settings.

Nevertheless, GeoSyn primarily modifies the magnitude and direction of estimated motion, whereas our method explicitly augments trajectory geometry and temporal motion dynamics. We will further clarify the scope and limitations of this comparison in the revised manuscript.


[A5] Generalization to Real-World Degradations

We appreciate the reviewer's concern regarding generalization to complex real-world degradations, such as rolling-shutter effects and sensor noise. These effects depend on camera-specific imaging processes and are difficult to explicitly model within our trajectory-based augmentation framework. Nevertheless, we can evaluate the generalization ability of our method on real-world datasets that reflect practical camera acquisition conditions. To this end, we consider both our existing RSBlur evaluation and additional cross-domain experiments on RSBlur and RealBlur-J/R.

As shown in Table 2 of the revised manuscript (reproduced below), our method achieves higher PSNR on RSBlur than the unaugmented baseline and ID-Blau under our training protocol, and also exceeds the result reported in the original GeoSyn paper. Furthermore, Table 5 of the revised manuscript (reproduced below) shows that, when trained on GoPro, our method generally achieves better performance than the unaugmented baseline and ID-Blau on RSBlur, RealBlur-J, and RealBlur-R.

MethodsGMACs↓PSNR↑
NAFNet63.6433.97
+ ID-Blau63.6434.01
+ GeoSyn63.6434.23
+ Ours63.6434.34
Test SetAug.PSNR↑LPIPS↓
GoPro (Train) → Real-world Blur (Test)
RSBlur–27.360.334
ID-Blau26.340.386
Ours27.840.310
RealBlur-J–26.540.174
ID-Blau26.350.190
Ours26.550.165
RealBlur-R–33.650.097
ID-Blau33.760.095
Ours33.770.080

[1] Exposure Trajectory Recovery from Motion Blur, TPAMI 2021.

Thank you for your detailed response. My concerns have been addressed, and I will therefore maintain my positive rating.

We sincerely appreciate your encouraging follow-up and continued support. Your valuable feedback has helped us strengthen our work by more thoroughly examining the benefits and computational costs of our framework and presenting experimental comparisons more transparently. We are pleased that our responses have addressed your concerns, and we have incorporated the additional results and clarifications into the revised manuscript. We are truly grateful for your careful review and constructive suggestions, which have been instrumental in improving our work.

Reviewer_RhmjOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a trajectory-aware differentiable blur augmentation framework. It employs Global Trajectory-aware Augmentation (GTA) and Local Trajectory-aware Augmentation (LTA) to model global motion and spatially varying local motion residuals, respectively, and synthesizes blurry images through a differentiable renderer. A learnable Augmentor is further introduced to adaptively perturb trajectory parameters and generate more challenging training samples. Experimental results demonstrate the effectiveness of the proposed method.

Critical strengths
  1. Rather than limiting blur augmentation to changes in motion magnitude and orientation, this paper introduces parameterized motion dynamics and trajectory primitives to further diversify the geometric and temporal variations of blur patterns. The combination of GTA/LTA and the differentiable Augmentor forms a relatively comprehensive and interpretable blur augmentation framework.
  2. The effectiveness of the proposed method is validated across multiple backbones, different training data sizes, and cross-dataset settings. The comparison with random perturbations further demonstrates the benefit of differentiable trajectory optimization.
Critical weaknesses
  1. Trajectory diversity is a key motivation of this work. However, Figure 1© relies on the authors’ own pretrained trajectory estimation network to analyze motion distributions before and after augmentation. Moreover, the clustering is based on normalized trajectory geometry and thus cannot directly reflect diversity in temporal motion dynamics. In addition, the decomposition of global and local motion in GTA/LTA lacks independent validation, making it difficult to determine the extent to which the learned global and local motion parameters correspond to actual camera and object motions.
Questions for authors
  1. Could the authors provide trajectory analysis or validation independent of the proposed Reblurring Network to further support the claim of expanded trajectory-space coverage? In particular, how can the authors verify that the proposed augmentation genuinely increases the diversity of temporal motion dynamics, rather than merely altering the geometric trajectory distributions predicted by their own estimator?
1 comment

We thank the reviewers for their constructive feedback and address their concerns below. Newly added text and figure/table captions in the revised manuscript are temporarily highlighted in blue\textcolor{blue}{\text{blue}}, while table contents remain black for readability. We have also condensed existing text to accommodate additional experiments and welcome further discussion.


[A1] Independent Validation of Trajectory Diversity and Motion Dynamics

We appreciate the reviewer's concerns regarding the independent validation of trajectory diversity and temporal motion dynamics. We agree that evaluating trajectories using our own estimator may introduce bias and that geometric trajectory analysis alone may not fully reflect temporal motion variations. To address these concerns, we conducted an additional analysis using Motion-ETR [1], an independent method that estimates motion trajectories over the exposure interval while accounting for their temporal evolution.

For Figure 1(c) of the manuscript, we resample each trajectory at uniform arc-length intervals, which removes information about the time taken to traverse different parts of the path. For Figure 7 of the revised manuscript, we retain the center-pixel trajectory estimated by Motion-ETR[1] at 15 uniformly spaced times during exposure, as described in Appendix B.4. The distance between successive points therefore retains information about relative speed along the estimated path. Under this representation, trajectories estimated from our augmented images are more broadly distributed across clusters than those estimated from ID-Blau images. Since the clusters reflect both trajectory shape and relative timing, this analysis does not separately quantify temporal variation.


[A2] Validation of Global and Local Motion Modeling

We appreciate the reviewer's concern regarding the validation of global and local motion modeling. GTA and LTA are designed to capture global and spatially varying local motion patterns for blur augmentation, rather than explicitly estimating ground-truth camera and object motion. We will clarify this distinction in the revised manuscript.

As ground-truth camera and object trajectories are unavailable in the datasets, we assess the contributions of GTA and LTA indirectly through ablation experiments. Specifically, we expanded our ablation study by including GTA-only augmentation with random trajectory perturbations. As shown in Table 3 of the revised manuscript (reproduced below), this enables a systematic comparison of GTA and GTA+LTA under both random and differentiable augmentation settings. The results demonstrate the complementary benefits of global and local trajectory modeling for effective blur augmentation.

GTALTADifferentiableGoProHIDE
MUSIQ↑LPIPS↓PSNR↑DISTS↓MUSIQ↑LPIPS↓PSNR↑DISTS↓
38.490.15528.920.09743.640.19226.710.113
✓42.780.10831.080.08049.940.15528.630.088
✓✓43.200.10431.240.07750.640.15028.690.085
✓✓42.390.10631.190.07750.760.14728.780.083
✓✓✓43.420.09931.370.07451.370.14228.980.081

[1] Exposure Trajectory Recovery from Motion Blur, TPAMI 2021.

Reviewer_zsyWOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a differentiable trajectory-aware augmentation framework for image deblurring. The Global Trajectory-aware Augmentation module represents camera motion using velocity, acceleration, curve and S-curve primitives, and in-plane rotation, while the Local Trajectory-aware Augmentation module uses spatially weighted motion components to model local motion. A pretrained reblurring network infers trajectory parameters from blur-sharp pairs, and a learnable Augmentor perturbs these parameters under the deblurring objective to generate challenging training samples. Experiments cover 10%, 50%, and 100% GoPro training settings, cross-dataset evaluation on HIDE, evaluation on RSBlur, and multiple deblurring backbones.

Critical strengths
  1. Augmenting structured motion trajectories rather than only blur magnitude and orientation is a clear and meaningful research direction.
  2. Improvements are generally consistent across MIMO-UNet+, Concertormer, and NAFNet, as well as across synthetic and real-blur benchmarks.
Critical weaknesses

Weaknesses and Questions

  1. Although the proposed trajectory-aware blur augmentation aims to improve deblurring generalization, the experiments lack cross-domain evaluation between synthetic and real-world blur. Specifically, GoPro-trained models are evaluated only on GoPro and HIDE, while RSBlur is used for separate training and testing. The authors should conduct synthetic-to-real (GoPro → RSBlur) and real-to-synthetic (RSBlur → GoPro/HIDE) evaluations, directly comparing the proposed blur augmentation with existing augmentation strategies (e.g., ID-Blau and random trajectory perturbations) under identical training settings. Such comparisons are important to determine whether the proposed trajectory-aware augmentation offers a genuine advantage over existing strategies in generalizing to unseen blur distributions across the synthetic–real domain gap.

  2. The experimental evaluation primarily focuses on CNN- and Transformer-based deblurring models. It would be beneficial to extend the comparisons to recent Mamba-based architectures [1,2], to better demonstrate the applicability of the proposed augmentation strategy across diverse network architectures.

  3. Unlike separately optimized blur augmentation methods, the proposed Augmentor is optimized using feedback from the target deblurring network, potentially introducing strong backbone dependency. It remains unclear whether the learned augmentation strategy can generalize across different deblurring architectures without retraining. Evaluating the transferability of an Augmentor trained with one backbone to other backbones would help clarify this concern.

  4. In Figure 1(c), the authors use their own trajectory estimator to evaluate the diversity of trajectories generated by different augmentation methods. Since this estimator is closely related to the proposed trajectory model, the evaluation may favor the proposed method. Using an independent trajectory estimator or synthetic data with ground-truth trajectories would make the comparison more convincing.

  5. The Augmentor is optimized to generate challenging blur patterns by maximizing the deblurring loss, but more challenging samples are not necessarily more realistic. Although the authors constrain trajectory perturbations, it remains unclear whether the generated blur patterns reflect realistic camera and object motion. Additional analysis of blur realism would help strengthen this claim.

  6. The component ablation in Table 3 lacks some important configurations, particularly GTA with random trajectory perturbations. Including this setting would help isolate the contribution of differentiable augmentation from trajectory modeling and clarify whether its effectiveness depends on the inclusion of LTA.

  7. The proposed LTA models local motion through weighted combinations of 2D trajectories without explicitly considering depth or occlusion. This may lead to unrealistic blur near moving object boundaries. Moreover, synthesizing blur by warping a single sharp image cannot accurately account for occlusion and disocclusion during exposure. Further analysis of these effects would help assess the physical realism of the generated blur.

[1] Kong et al., "Efficient Visual State Space Model for Image Deblurring", CVPR 2025.

[2] Li et al., "MaIR: A Locality- and Continuity-Preserving Mamba for Image Restoration", CVPR 2025.

Questions for authors

Please see the weaknesses.

6 comments

We thank the reviewers for their constructive feedback and address their concerns below. Newly added text and figure/table captions in the revised manuscript are temporarily highlighted in blue\textcolor{blue}{\text{blue}}, while table contents remain black for readability. We have also condensed existing text to accommodate additional experiments and welcome further discussion.


[A1] Cross-domain Generalization

We agree that cross-domain evaluation is important for assessing the generalization ability of our method. To further address this concern, we conducted additional experiments using Concertormer [1], evaluating GoPro →\rightarrow RSBlur/RealBlur and RSBlur →\rightarrow GoPro/HIDE under consistent experimental settings, comparing our method with the baseline and ID-Blau, which employs random perturbations.

As shown in Table 5 of the revised manuscript (reproduced below), our method generally outperforms the baseline and ID-Blau across both domains. These results further demonstrate the effectiveness of our trajectory-aware augmentation in generalizing to unseen blur distributions across the synthetic–real domain gap. The ID-Blau comparisons in the RSBlur → GoPro/HIDE setting are still in progress, and we will report the results during the discussion period if available, or once completed.

Test SetAug.PSNR↑LPIPS↓
GoPro (Train) → Real-world Blur (Test)
RSBlur–27.360.334
ID-Blau26.340.386
Ours27.840.310
RealBlur-J–26.540.174
ID-Blau26.350.190
Ours26.550.165
RealBlur-R–33.650.097
ID-Blau33.760.095
Ours33.770.080
RSBlur (Train) → Synthetic Blur (Test)
GoPro–26.470.261
Ours27.870.260
HIDE–25.130.172
Ours26.090.182

[A2] Evaluation on Mamba-based Architectures

We appreciate the reviewer's valuable suggestion to extend our evaluation to Mamba-based architectures. To address this point, we conducted additional experiments using EVSSM [2], a Mamba-based deblurring network. As shown in Table 7 of the revised manuscript (reproduced below), our method outperforms the baseline and ID-Blau, further demonstrating its effectiveness across different network architectures.

DatasetAug.MUSIQ↑PSNR↑LPIPS↓DISTS↓
GoPro–37.8930.060.1430.092
ID-Blau42.3031.350.1130.083
Ours42.9831.820.1020.077
HIDE–45.0328.460.1660.098
ID-Blau50.1229.550.1370.084
Ours51.4130.040.1280.078

[A3] Cross-backbone Transferability of the Augmentor

We appreciate the reviewer's concern regarding the backbone dependency of our Augmentor. A key aspect of our approach is that the Augmentor is optimized online using feedback from the deblurring network, allowing the augmentation strategy to adapt to each backbone during training. Transferring a pretrained Augmentor without further optimization removes this adaptive mechanism and may therefore lead to performance degradation. To examine this effect, we conduct cross-backbone transfer experiments between Concertormer and MIMO-UNet+ in both directions, comparing online optimization with each backbone (Specific) against frozen transfer without further optimization (Transfer).

As shown in Table 6 of the revised manuscript (reproduced below), the transferred Augmentor generally shows favorable performance compared to ID-Blau across both backbones, even without online optimization. This suggests that our learned trajectory-aware augmentation remains effective across different architectures. Meanwhile, backbone-specific online optimization achieves further improvements over frozen transfer, highlighting the benefit of adapting trajectory perturbations to the current deblurring objective. Together, these results support both the effectiveness of our trajectory-aware motion modeling and the importance of feedback-driven online optimization, without requiring backbone-specific modifications to the Augmentor architecture.

BackboneAug.GoPro PSNR↑GoPro LPIPS↓HIDE PSNR↑HIDE LPIPS↓
Concertormer–30.470.13628.820.160
ID-Blau30.520.13328.770.158
Ours (Transfer)30.620.13229.000.157
Ours (Specific)31.390.11629.610.148
MIMO-UNet+–28.920.15526.710.192
ID-Blau30.640.12328.500.157
Ours (Transfer)30.700.11428.360.169
Ours (Specific)31.370.09928.980.142

[A4] Independent Evaluation of Trajectory Diversity

We acknowledge the reviewer's concern that evaluating trajectory diversity using our own estimator may introduce bias. To address this concern, we conduct an additional analysis using Motion-ETR[3], an independent trajectory estimator.

To independently evaluate trajectory diversity, we use Motion-ETR to estimate center-pixel trajectories at 15 uniformly spaced exposure times, as described in Appendix B.4. This representation preserves relative temporal information along each trajectory. As shown in Figure 7 of the revised manuscript, our method exhibits broader trajectory-space coverage than ID-Blau, consistent with the findings in Figure 1(c) of the revised manuscript. These results further support the diversity of our generated motion patterns using an independent trajectory estimator.


[A5] Physical Realism of Generated Blur

We agree with the reviewer that generating more challenging blur patterns does not necessarily guarantee their realism. Our primary objective is to improve downstream deblurring performance by adaptively generating blur patterns tailored to the training dataset and target network. To this end, our method aims to expand the diversity of existing blur patterns by introducing a wider range of motion trajectories. While we acknowledge that our approach cannot guarantee the physical realism of generated blur, we aim to increase blur diversity without deviating excessively from realistic blur distributions.

To examine this, we evaluate recall and coverage for diversity, along with FID and CMMD for distributional similarity. As shown in Table 4 of the revised manuscript (reproduced below), our method achieves higher recall and coverage than ID-Blau, while also obtaining lower FID and CMMD against the original GoPro blur distribution. Together with the broader trajectory diversity shown in Figures 1(c) and 7 of the revised manuscript, these results suggest that, although physical realism cannot be directly verified, our method generates diverse blur patterns without deviating substantially from the original GoPro blur distribution.

MethodFID↓CMMD↓Recall↑Coverage↑
ID-Blau26.490.0940.9440.978
Ours25.960.0350.9700.984

[A6] Additional Ablation Studies

We thank the reviewer for identifying the missing ablation setting. We have added a GTA-only experiment with random trajectory perturbations, enabling a direct comparison of random and differentiable augmentation with and without LTA, as shown in Table 3 of the revised manuscript (reproduced below).

GTALTADifferentiableGoProHIDE
MUSIQ↑LPIPS↓PSNR↑DISTS↓MUSIQ↑LPIPS↓PSNR↑DISTS↓
38.490.15528.920.09743.640.19226.710.113
✓42.780.10831.080.08049.940.15528.630.088
✓✓43.200.10431.240.07750.640.15028.690.085
✓✓42.390.10631.190.07750.760.14728.780.083
✓✓✓43.420.09931.370.07451.370.14228.980.081

[A7] Depth and Occlusion Limitations of LTA

We agree with the reviewer that our LTA does not explicitly model depth and occlusion effects. In particular, spatially weighted blending of 2D trajectories may produce inaccurate blur near object boundaries where foreground and background motions differ. As discussed in prior image formation studies [4, 5], accurately modeling motion blur in such regions requires accounting for time-varying visibility and foreground-background interactions during exposure. Since previously occluded scene content cannot be uniquely inferred from a single sharp image, single-image-based blur synthesis cannot, in general, fully reproduce physically accurate occlusion and disocclusion effects without additional scene information or assumptions.

Fully addressing these effects would require more comprehensive scene and visibility modeling beyond 2D trajectory-based warping. Our work instead focuses on trajectory-aware blur augmentation through GTA and LTA to approximate global and spatially varying motion patterns. We will clarify this limitation in the revised manuscript and consider explicit depth and occlusion modeling in future work.

[1] Efficient Concertormer for Image Deblurring and Beyond, ICCV 2025.

[2] Efficient Visual State Space Model for Image Deblurring, CVPR 2025.

[3] Exposure Trajectory Recovery from Motion Blur, TPAMI 2021.

[4] Motion Field Estimation from Alternate Exposure Images, TPAMI 2011.

[5] Removing Partial Blur in a Single Image, CVPR 2009.

We sincerely thank you for your thoughtful review and valuable constructive feedback. We appreciate the time and effort you have devoted to evaluating our work. Your insightful comments have been very helpful in strengthening our work through additional experiments and clearer discussions of blur diversity, realism, and limitations. We have carefully incorporated these improvements into our rebuttal responses.

We hope that our responses have adequately addressed your concerns and would appreciate your thoughts on whether any concerns remain. If further clarification is needed, we would be happy to address it during the remaining discussion period. Thank you again for your valuable time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.