acceptodds
Withdrawn from ICLR 2027

IMD-Blur: Implicit Motion and Defocus Blur Augmentation for Image Deblurring

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The scarcity of paired sharp/blur training data for image deblurring makes realistic blur synthesis important for data augmentation and model generalization. However, existing blur augmentation methods typically focus on isolated blur types, such as defocus or motion blur, and often rely on auxiliary signals, including optical flow, aperture settings, or depth/defocus maps, which are rarely available in practice. To address these limitations, we propose IMD-Blur (Implicit Motion and Defocus Blur Augmentation), a unified framework that learns pixel-wise blur parameters directly from paired sharp/blur images without external signals. Specifically, using a parameterized kernel-based blurring pipeline and a blur parameter estimator, IMD-Blur constructs a continuous and interpretable parameter space for observed blur patterns. Based on the estimated blur parameters, we design a diffusion-based blur generator that produces diverse and controllable motion- and defocus-blurred images, thereby densifying the observed blur distribution and improving training-distribution coverage. Experiments show that IMD-Blur produces realistic blurred images and consistently improves state-of-the-art deblurring models across multiple benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Deblurring, Low-level Vision, Image Restoration

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

This paper proposes IMD-Blur, a data augmentation framework for image deblurring that learns to synthesize motion and defocus blur using paired sharp/blur images without requiring auxiliary annotations such as optical flow or depth. The framework consists of two stages. First, PKBP estimates pixel-wise blur parameters from paired images and maps them to differentiable blur kernels. Motion blur is represented using magnitude, orientation, and curvature, while defocus blur is represented using a circle-of-confusion radius. Second, DBG refines the kernel-blurred images to reproduce domain-specific appearance effects. For augmentation, blur parameters are sampled from distributions estimated on the training data, and the resulting synthetic pairs are used to pretrain deblurring networks before task-specific fine-tuning.

Critical strengths
  1. Practical problem formulation: Generating realistic blur augmentation without requiring optical flow, depth maps, or camera metadata is practically valuable. Such auxiliary information may be unavailable or unreliable for real-world deblurring datasets. Learning blur synthesis directly from paired sharp/blur images makes the proposed approach applicable to a broader range of existing training datasets.

  2. Interpretable design: The framework separates parameterized blur synthesis from learned appearance refinement. PKBP provides explicit control over motion magnitude, direction, curvature, and defocus radius, while DBG aims to model effects that are difficult to capture using simple analytical kernels. This separation is conceptually appealing and enables component-level analysis.

  3. Broad evaluation across blur types and deblurring architectures: The experiments cover both motion and defocus blur, multiple deblurring architectures, and several datasets. Evaluation on HIDE, RealDOF, and RTF also provides evidence of cross-dataset generalization, rather than restricting the results to the training domains.

Critical weaknesses
  1. The relationship between pixel-wise blur estimation and the actual augmentation procedure is unclear: A central contribution is the estimation and synthesis of spatially varying blur. However, Appendix A states that during data augmentation, motion parameters and defocus radius are sampled once per image and applied uniformly across all pixels. This raises an important distinction between the proposed pixel-wise blur representation and the blur distributions actually used to generate augmented training data.

  2. The claimed improvement in blur-distribution coverage is not directly demonstrated: The paper motivates IMD-Blur as a way to densify the observed blur distribution and improve training-distribution coverage. However, the augmentation parameters are sampled from empirical distributions estimated on the training set. For example, motion blur caused by camera translation along the optical (depth) axis produces spatially varying radial motion patterns that may not be adequately captured by the proposed parameterization and sampling strategy.

  3. The novelty: The proposed framework is related to prior conditional blur-synthesis methods, particularly ID-Blau. Its main distinctions appear to be learning parameterized blur conditions directly from paired images, introducing explicit analytical kernels, and refining the resulting images using a conditional diffusion generator. These are potentially useful contributions, but the paper provides limited analysis of how much each distinction matters relative to simpler or existing alternatives.

Questions for authors
  1. Distribution coverage: Can the authors quantitatively demonstrate that the generated data improve coverage of real or held-out blur distributions? How are correlations among motion parameters handled during sampling?

  2. Computational cost: What is the total training cost of PKBP, DBG, augmentation generation, and deblurring-network training?

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

This paper proposes IMD-Blur, a framework for controllable motion and defocus blur augmentation using paired sharp and blurred images without requiring external auxiliary signals. The proposed method consists of two main components: a Parameterized Kernel-based Blurring Pipeline (PKBP) and a Diffusion-based Blur Generator (DBG). PKBP implicitly estimates interpretable blur parameters through reconstruction-based supervision, representing motion blur using magnitude, orientation, and curvature, and defocus blur using the circle-of-confusion radius. These estimated parameters are subsequently used to construct spatially varying blur kernels. DBG further utilizes the estimated parameters and kernel-based blurred images as conditions to synthesize realistic, domain-specific blurred images. A distribution-aware sampling strategy is introduced to generate additional training samples for deblurring models. Experiments on multiple motion and defocus deblurring benchmarks demonstrate consistent improvements across different network architectures, supported by comparative evaluations and component-level ablation studies.

Critical strengths

The paper addresses an important problem in image deblurring by proposing a controllable blur augmentation framework that supports both motion and defocus blur without requiring auxiliary supervision. The combination of interpretable blur parameter estimation and diffusion-based generation is well motivated, providing a flexible approach to synthesizing diverse blur patterns. In particular, introducing curvature into the motion blur parameterization enables the modeling of non-linear motion trajectories beyond conventional magnitude- and orientation-based representations. The proposed framework also demonstrates consistent performance improvements across multiple deblurring architectures and benchmarks, including GoPro, HIDE, RealBlur, and DPDD. Furthermore, the comparisons with existing blur synthesis methods, component-level ablations, and additional experiments under matched training budgets provide useful empirical evidence for the effectiveness of the proposed approach.

Critical weaknesses

Despite the consistent performance improvements, several limitations of the proposed framework warrant further investigation. First, although IMD-Blur is presented as a unified framework for motion and defocus blur augmentation, the two blur types are handled through separate parameter estimation branches, and the blur type is assumed to be known based on the data acquisition protocol. Consequently, the framework does not explicitly demonstrate its ability to model images containing simultaneous motion and defocus blur, limiting the strength of its unified modeling claim. Second, while the paper emphasizes pixel-wise blur parameter estimation and spatially varying blur synthesis, Appendix A states that the parameters used for actual data augmentation are sampled once per image and applied uniformly across all pixels. This raises questions about the extent to which the reported improvements can be attributed to spatially varying blur generation. Furthermore, the proposed distribution-aware sampling strategy relies on estimated blur parameter distributions, but it remains unclear whether the correlations among different blur parameters and their spatial dependencies are adequately preserved. Since the parameters are learned solely through reconstruction supervision, their physical interpretability and consistency with real blur formation also require stronger validation. In addition, the experimental improvements over competing augmentation methods are sometimes relatively modest, and the contribution of each modeling choice to the overall performance could be more thoroughly investigated. Finally, although the appendix reports blur synthesis inference costs, the computational expense of the complete augmentation and training pipeline remains insufficiently characterized, particularly considering the additional diffusion model training and generation stages.

Questions for authors
  1. Can the authors clarify how the proposed unified framework handles images containing simultaneous motion and defocus blur, given that the two blur types are modeled separately and assumed to be known?

  2. Appendix A states that blur parameters are sampled once per image and applied uniformly during augmentation, whereas Appendix E emphasizes the benefits of spatially varying blur estimation. Could the authors clarify how spatially varying information is preserved in the actual augmentation process and reconcile these descriptions?

  3. How does the proposed distribution-aware sampling strategy preserve the correlations among different blur parameters and their spatial dependencies to ensure physically plausible blur generation?

  4. Can the authors provide additional quantitative evidence that the implicitly estimated motion and defocus parameters correspond to physically meaningful blur characteristics rather than merely minimizing reconstruction errors?

  5. Can the authors further clarify the experimental protocol underlying the different GoPro results reported in Tables 4 and 6, and provide a more comprehensive analysis of the computational cost of the complete training and augmentation pipeline?

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

This paper proposes IMD-Blur, a blur augmentation framework combining parameterized kernels with a conditional diffusion generator. Motion blur is represented by pixel-wise magnitude, orientation, and curvature, while defocus blur is represented by a circle-of-confusion radius. A parameter estimator learns these latent blur fields from sharp-blur pairs through reconstruction, and a Diffusion-based Blur Generator is trained to add domain-specific appearance effects not captured by the analytical kernels. The authors generate an augmented dataset five times larger than the original data and report improvements across several motion- and defocus-deblurring models and benchmarks.

Critical strengths
  1. The overall framework applies a common high-level idea to both motion and defocus blur, giving it broader scope than methods designed for only one degradation type.

  2. The experimental scope is broad, covering several backbones and datasets such as GoPro, HIDE, RealBlur, DPDD, RealDOF, and RTF.

Critical weaknesses

Weaknesses and Questions

  1. The claim of a unified motion-and-defocus framework is overstated. Blur type is assumed to be known from the acquisition protocol; motion and defocus use different estimators and kernel mappings and are trained separately on GoPro and DPDD. The paper does not demonstrate a single model that recognizes both blur types or handles images containing mixed motion and defocus blur. The implementation is therefore better described as two parallel pipelines sharing a general design than as a unified mixed-blur model.

  2. The main paper emphasizes pixel-wise controllable and spatially varying blur generation, whereas Appendix A states that motion parameters (magnitude, orientation, and curvature) and defocus parameters (CoC radius) are each sampled once per image and applied uniformly across all pixels during augmentation. This raises questions about whether the actual augmentation process fully exploits the proposed spatially varying representation. Furthermore, Table 10 compares spatially uniform and pixel-wise parameter estimation rather than directly evaluating different parameter sampling strategies. A controlled comparison between spatially uniform and spatially varying augmentation would help clarify the contribution of the proposed pixel-wise modeling.

  3. Although DBG is designed to generate controllable blur conditioned on the estimated parameters, the experiments do not sufficiently establish whether it actually relies on their intended semantics. Ablations removing (P) or (B_k), as well as quantitative evaluations of parameter sensitivity, would help clarify their respective contributions. Furthermore, Table 7 evaluates reconstruction fidelity against target blurred images, but it is unclear how well this reflects the generation of novel blur patterns from independently sampled parameters. Since the parameter estimator takes the target blurred image as input, the reconstruction evaluation may also benefit from target-specific information encoded in the estimated parameter fields.

  4. The blur parameters are learned implicitly through reconstruction losses, without direct evaluation against ground-truth motion trajectories or defocus parameters. Although the downstream improvements demonstrate their usefulness for augmentation, they do not necessarily establish the physical accuracy or interpretability of the estimated parameters. Additional experiments using synthetic blur with known parameters would strengthen this claim.

  5. While the ablation studies demonstrate improvements from the Diffusion-based Blur Generator, it remains unclear whether these gains specifically arise from diffusion modeling or simply from introducing an additional learnable refinement network. Comparisons with simpler conditional generation or residual refinement networks under comparable model capacities would better justify the use of diffusion.

Questions for authors

Please see weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.