Efficient Real-World Deblurring using Single Images: AIM 2025 Challenge Report

TL;DR

The paper surveys the AIM 2025 Efficient Real-World Deblurring using Single Images challenge, which targets real-world motion blur restoration under strict efficiency constraints on RSBlur-based data captured with a dual-camera system. It documents four top-performing approaches that adapt and optimize NAFNet and Restormer architectures for low-parameter, low-GMAC regimes, achieving a peak PSNR around 31.13 dB while meeting the 5M-parameter and 200 GMACs limits. The study highlights the robustness of efficiency-focused design, including progressive patch training, data augmentation for generalization, and spatial-attention enhancements, demonstrated on real blurred images rather than synthetic benchmarks. Practically, the work provides baselines and methodological guidance for deploying efficient deblurring on mobile/edge devices, and it outlines future directions for more challenging, metadata-enabled real-world inverse problems.

Abstract

This paper reviews the AIM 2025 Efficient Real-World Deblurring using Single Images Challenge, which aims to advance in efficient real-blur restoration. The challenge is based on a new test set based on the well known RSBlur dataset. Pairs of blur and degraded images in this dataset are captured using a double-camera system. Participant were tasked with developing solutions to effectively deblur these type of images while fulfilling strict efficiency constraints: fewer than 5 million model parameters and a computational budget under 200 GMACs. A total of 71 participants registered, with 4 teams finally submitting valid solutions. The top-performing approach achieved a PSNR of 31.1298 dB, showcasing the potential of efficient methods in this domain. This paper provides a comprehensive overview of the challenge, compares the proposed solutions, and serves as a valuable reference for researchers in efficient real-world image deblurring.

Figures (7)

• Figure 1: A dual-camera system for collecting the RSBlur dataset.
• Figure 2: Sample of the testset used for the challenge.
• Figure 3: The architecture of NAFRepLocal proposed by MiVideoDeblur.
• Figure 4: Overall architecture of RestormerL. The model employs a symmetric encoder-decoder design with skip connections and utilizes Multi-Dconv Head Transposed Attention (MDTA) and Gated-Dconv Feed-Forward Network (GDFN) blocks. Downsampling and upsampling are achieved via pixel-unshuffle and pixel-shuffle operations, respectively.
• Figure 5: The Overview the proposed method architecture by Team IPIU.