DRAMNet: Depth-initialized Region-Adaptive Map Network for Single-Image Deblurring

ICLR 2026 Conference Submission14408 Authors

Keywords: deblurring, depth, attention, pretraining, fusion, priors, architecture, dual-branch

TL;DR: DRAMNet leverages a frozen depth‐estimation head, a multi‐scale blur‐map branch, and a reversible deblurring decoder fused via cross‐attention to achieve state‐of‐the‐art single‐image deblurring

Abstract: Recent advances in image deblurring have achieved impressive results, yet existing methods still struggle with two key challenges: the scarcity of training data compared to other image restoration tasks and the inability to effectively handle variable blur strength across different image regions. We present DRAMNet, a three-part system that addresses these issues by transferring knowledge from the depth estimation task and using a specially designed component to assess and adapt to varying blur strength across the image. Per-patch blur map estimation allows the model to react differently to heavily and lightly blurred sub-regions, while depth information from pre-training provides structural guidance even with limited deblurring-specific data. Extensive experiments on the most popular synthetic (GoPro, REDS) and real-world (RSBlur, RealBlur) benchmarks show that DRAMNet outperforms state-of-the-art methods across the PSNR, SSIM, and LPIPS metrics. We made our code available at [Link is removed for blind review].

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 14408