Reproduce “Efficient Real-World Deblurring using Single Images: AIM 2025 Challenge Report” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Daniel Feijoo, Paula Garrido-Mellado, Marcos V. Conde, Jaesung Rim, Alvaro Garcia, Sunghyun Cho, Radu Timofte
The AIM 2025 challenge report details a competition focused on efficient real-world single-image deblurring, demonstrating that high-quality image restoration is achievable within strict computational budgets using real-world blur data. The winning solution, NAFRepLocal, achieved a PSNR of 31.130 dB and SSIM of 0.843 while adhering to limits of fewer than 5 million parameters and under 200 GMACs.
Key insights
- High-quality real-world image deblurring can be achieved under strict computational budgets (e.g., <5M parameters, <200 GMACs), making on-device deployment feasible.
- Architectures like NAFNet and streamlined Transformers (Restormer) serve as strong foundations for developing efficient deblurring models when combined with careful modifications.
- Multi-stage progressive training, reparameterization, localized attention, and simplified components are effective strategies for optimizing both performance and efficiency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.