Event-Conditioned Dual-Modal Fusion for Motion Deblurring

Motion deblurring aims to recover sharp frames from blurred inputs caused by camera shake or object movement during exposure. While deep learning-based methods have shown promising results, they often struggle with complex motion due to the absence of temporal cues in a single frame. Event cameras, with their high temporal resolution, provide complementary signals that capture finegrained motion information, making them particularly wellsuited for addressing motion blur. However, existing approaches typically focus on either the spatial or temporal aspects of event data in isolation, limiting their ability to fully exploit the rich information events offer. Additionally, these models often fail to generalize to realworld scenarios where motion patterns are diverse and blur severity varies. To address these challenges, we propose two complementary models: EFSformer, which enhances spatial structure through a novel event-image fusion block operating in both temporal and frequency domains, and Video-SFHformer, which captures long-range motion using a lightweight spatio-temporal extraction block. We further introduce an ensemble strategy to integrate the strengths of both models, resulting in a unified framework named EV-Deblurformer. Extensive experiments on realworld datasets demonstrate the superiority of our method over existing state-of-the-art approaches. Notably, EVDeblurformer achieved the highest SSIM score among all submissions and ranked third overall in the NTIRE 2025 Event-Based Image Deblurring Challenge.