EventSR-Zero: Training-free Event Video Super-Resolution with Diffusion Priors

Jie Long Lee, Gim Hee Lee

Keywords: event cameras, super resolution, diffusion

TL;DR: EventSR-Zero leverages the high temporal resolution of event cameras to recover fine details from low-resolution events and guides a diffusion-based Video Super-Resolution model to generate high-quality, super-resolved videos of the scene

Abstract: Event-to-Video (E2V) methods aim to reconstruct intensity frames from events, bridging the gap between event-based and image-based vision. However, existing E2V approaches often fail to recover fine structures, leading to reconstructions with artifacts and degraded quality. To address this, we explore the task of Event-to-Video Super-Resolution (EVSR), which aims to reconstruct high-resolution video from low-resolution events. We present EventSR-Zero, a training-free framework that exploits the high temporal resolution of event cameras to recover fine-grained details from low-resolution events and uses them to guide a diffusion-based Video Super-Resolution (VSR) model in generating high-quality super-resolved videos of the underlying scene. Our approach incorporates two key components: (1) an Implicit Contrast Refinement (ICR) module that robustly extracts sub-pixel scene details from low-resolution events, and (2) a Reconditioning Guidance (RG) module that reliably steers the diffusion VSR process using the high-resolution event signal from ICR. Extensive experiments demonstrate that EventSR-Zero achieves state-of-the-art performance, surpassing existing event-based super-resolution methods. We will release our source code upon acceptance.

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

Submission Number: 7268