Abstract
In real-world scenarios, image impairments often manifest as compositedegradations, presenting a complex interplay of elements such as low light,haze, rain, and snow. Despite this reality, existing restoration methodstypically target isolated degradation types, thereby falling short inenvironments where multiple degrading factors coexist. To bridge this gap, ourstudy proposes a versatile imaging model that consolidates four physicalcorruption paradigms to accurately represent complex, composite degradationscenarios. In this context, we propose OneRestore, a novel transformer-basedframework designed for adaptive, controllable scene restoration. The proposedframework leverages a unique cross-attention mechanism, merging degraded scenedescriptors with image features, allowing for nuanced restoration. Our modelallows versatile input scene descriptors, ranging from manual text embeddingsto automatic extractions based on visual attributes. Our methodology is furtherenhanced through a composite degradation restoration loss, using extra degradedimages as negative samples to fortify model constraints. Comparative results onsynthetic and real-world datasets demonstrate OneRestore as a superiorsolution, significantly advancing the state-of-the-art in addressing complex,composite degradations.