WildRelight: A Real-World Dataset and Benchmark for Single-Image Relighting

Lezhong Wang, Mehmet Onurcan Kaya, Siavash Arjomand Bigdeli, Jeppe Revall Frisvad; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings, 2026, pp. 2007-2016

Recent single-image relighting methods, powered by advanced generative models, have achieved impressive photorealism on synthetic benchmarks. However, their effectiveness in the complex visual landscape of the real world remains largely unverified. A critical gap exists, as current real-world datasets are typically designed for multi-view reconstruction and fail to address the unique challenges of single-image relighting. To bridge this synthetic-to-real gap, we introduce WildRelight, the first in-the-wild dataset specifically created for evaluating and training single-image relighting models. WildRelight features a diverse collection of high resolution outdoor scenes, each paired with a spatially aligned, high-dynamic-range environment map that serves as the ground truth illumination. Our meticulous capture process, using a custom rig to precisely co-locate the nodal points of a primary camera and a 360 degree camera, ensures the physical accuracy of the lighting data. Using our dataset, we conduct the first comprehensive benchmark of state of the art methods. Our results quantify a significant domain gap, revealing poor zero shot performance for models trained only on synthetic data. Furthermore, we demonstrate that finetuning on WildRelight substantially improves model performance, validating its effectiveness for domain adaptation. The dataset and benchmark will be made publicly available to foster the next wave of research in robust, physically-grounded relighting.