When enhancing low-light images, many deep learning algorithms are based onthe Retinex theory. However, the Retinex model does not consider thecorruptions hidden in the dark or introduced by the light-up process. Besides,these methods usually require a tedious multi-stage training pipeline and relyon convolutional neural networks, showing limitations in capturing long-rangedependencies. In this paper, we formulate a simple yet principled One-stageRetinex-based Framework (ORF). ORF first estimates the illumination informationto light up the low-light image and then restores the corruption to produce theenhanced image. We design an Illumination-Guided Transformer (IGT) thatutilizes illumination representations to direct the modeling of non-localinteractions of regions with different lighting conditions. By plugging IGTinto ORF, we obtain our algorithm, Retinexformer. Comprehensive quantitativeand qualitative experiments demonstrate that our Retinexformer significantlyoutperforms state-of-the-art methods on thirteen benchmarks. The user study andapplication on low-light object detection also reveal the latent practicalvalues of our method. Code, models, and results are available athttps://github.com/caiyuanhao1998/Retinexformer