Beyond the Limits: Tackling Extreme Overexposure with Diffusion Model

In real-world photography scenarios, images suffer from overexposure due to overly bright environmental lighting conditions or incorrect camera settings. More severely, excessive illumination can cause extreme overexposure, where the brightness range exceeds the sensor’s recording capabilities and leads to loss of details in bright regions. We define these regions as brightness-saturated regions. We articulate the challenge of extremely overexposed image restoration in two aspects: firstly, the enhancement of image brightness and details; and secondly, restoring the information in the missing regions (i.e., the brightness-saturated regions). To address this issue, we propose a novel framework for extremely overexposed image restoration. Initially, we introduce an enhancement network aimed at adjusting the image’s brightness and contrast, striving to normalize the image’s brightness within the range of normal illumination. For the brightness-saturated regions, we first design a brightness extraction module to accurately extract them, and then we devise a semantic-guided large-model inpainting mechanism. Under the guidance of semantic information, we employ stable diffusion for information inpainting within these brightness-saturated regions, ensuring semantic consistency while maximally restoring the image. Compared to existing state-of-the-art methods, SEED achieves the best results on publicly available datasets.