UHD Image Deblurring via Autoregressive Flow with Ill-conditioned Constraints

Yucheng Xin, Dawei Zhao, Xiang Chen, Chen Wu, Pu Wang, Dianjie Lu, Guijuan Zhang, Xiuyi Jia, Zhuoran Zheng

Abstract

Ultra-high-definition (UHD) image deblurring poses significant challenges for UHD restoration methods, which must balance fine-grained detail recovery and practical inference efficiency. Although prominent discriminative and generative methods have achieved remarkable results, a trade-off persists between computational cost and the ability to generate fine-grained detail for UHD image deblurring tasks. To further alleviate these issues, we propose a novel autoregressive flow method for UHD image deblurring with an ill-conditioned constraint. Our core idea is to decompose UHD restoration into a progressive, coarse-to-fine process: at each scale, the sharp estimate is formed by upsampling the previous-scale result and adding a current-scale residual, enabling stable, stage-wise refinement from low to high resolution. We further introduce Flow Matching to model residual generation as a conditional vector field and perform few-step ODE sampling with efficient Euler/Heun solvers, enriching details while keeping inference affordable. Since multi-step generation at UHD can be numerically unstable, we propose an ill-conditioning suppression scheme by imposing condition-number regularization on a feature-induced attention matrix, improving convergence and cross-scale consistency. Our method demonstrates promising performance on blurred images at 4K (3840x2160) or higher resolutions.

Introduction

As mobile devices and professional cameras move toward Ultra-High-Definition (UHD) standards, image restoration tasks like deblurring face a "resolution dilemma." 4K images contain millions of pixels, causing standard Transformer-based models to hit memory limits or become prohibitively slow due to global attention mechanisms. Furthermore, while generative models like Diffusion excel at detail synthesis, they often become numerically unstable at UHD resolutions, leading to visual artifacts. Finally, most high-performing models are too heavy for real-world deployment on consumer hardware or mobile devices. The proposed ARF-IC (Autoregressive Flow with Ill-conditioned Constraints) framework aims to bridge this gap by creating a framework that is both high-quality and practically efficient.

Methodology

The ARF-IC framework relies on three main pillars to achieve efficient UHD deblurring:

  1. Autoregressive Coarse-to-Fine Architecture: Instead of deblurring the whole 4K image at once, the model works across multiple scales. It starts at a low resolution and progressively adds "residuals" (missing details) to upsampled versions of the previous scale. This ensures the model only learns what is newly resolvable at each step.
  2. Rectified Flow (Flow Matching): To generate these residuals, the authors use Flow Matching. Unlike standard diffusion models requiring hundreds of steps, Rectified Flow learns a "straight-line" transport path from noise to the target, allowing for high-quality results in just a few ODE sampling steps.
  3. Condition-Number Regularization: Generative processes at high resolutions are prone to error amplification. The authors characterize this instability via the condition number of the feature-induced attention matrix. By penalizing high condition numbers during training, they force numerical stability, preventing tiny errors from manifesting as artifacts.
Overview of the ARF-IC framework, highlighting the autoregressive architecture, flow matching, and condition-number regularization.

Experimental Setup

The model was evaluated on a mix of UHD-specific and standard-resolution benchmarks:

  • UHD-Blur: A synthetic motion-blur dataset at 4K resolution and higher.
  • MC-Blur (UHDM): A large-scale UHD dataset featuring complex motion blur and large kernel degradations.
  • Standard Benchmarks: GoPro, DVD, RealBlur-J, and RealBlur-R were used to test generalization to lower resolutions.

Results and Discussion

ARF-IC achieves state-of-the-art performance while maintaining high efficiency.

  • Performance: ARF-IC achieved the highest PSNR on the UHD-Blur (30.838 dB) and MC-Blur (28.328 dB) benchmarks, outperforming established models like Restormer and UHDformer.
  • Efficiency: On an RTX 3090, the model is significantly faster than many competitors. It can render a 4K image in under 2 seconds on modern mobile phones (Apple/Huawei).
  • Stability: Ablation studies showed that removing condition-number regularization caused a significant drop in PSNR (from 30.84 to 29.62 dB), proving that managing numerical stability is critical for UHD generative tasks.
Performance comparison on UHD-Blur and MC-Blur datasets. ARF-IC achieves superior PSNR with significantly lower runtime compared to state-of-the-art methods.

Conclusion and Future Work

While ARF-IC is explicitly optimized for UHD, its multi-scale overhead might be less efficient for very small images where single-pass networks suffice. Additionally, the "Resolution Control Strategy" can sometimes amplify sensor noise in extremely grainy inputs. Future work includes exploring this autoregressive flow approach for video deblurring and other UHD restoration tasks like deraining.

Summary updated

Abstract

Ultra-high-definition (UHD) image deblurring poses significant challenges for UHD restoration methods, which must balance fine-grained detail recovery and practical inference efficiency. Although prominent discriminative and generative methods have achieved remarkable results, a trade-off persists between computational cost and the ability to generate fine-grained detail for UHD image deblurring tasks. To further alleviate these issues, we propose a novel autoregressive flow method for UHD image deblurring with an ill-conditioned constraint. Our core idea is to decompose UHD restoration into a progressive, coarse-to-fine process: at each scale, the sharp estimate is formed by upsampling the previous-scale result and adding a current-scale residual, enabling stable, stage-wise refinement from low to high resolution. We further introduce Flow Matching to model residual generation as a conditional vector field and perform few-step ODE sampling with efficient Euler/Heun solvers, enriching details while keeping inference affordable. Since multi-step generation at UHD can be numerically unstable, we propose an ill-conditioning suppression scheme by imposing condition-number regularization on a feature-induced attention matrix, improving convergence and cross-scale consistency. Our method demonstrates promising performance on blurred images at 4K (3840$\times$2160) or higher resolutions.
BibTeX
@misc{xin2026uhd,
  author = {Yucheng Xin and Dawei Zhao and Xiang Chen and Chen Wu and Pu Wang and Dianjie Lu and Guijuan Zhang and Xiuyi Jia and Zhuoran Zheng},
  title = {{UHD Image Deblurring via Autoregressive Flow with Ill-conditioned Constraints}},
  year = {2026},
  eprint = {2603.10517},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.10517},
}
Figure