Existing lighting editing methods fail to simultaneously provide customizable lighting effect control and maintain content integrity. In this paper, we propose TransLight, a novel framework that enables high-fidelity and high-degree-of-freedom lighting effect transfer. The key is to extract lighting effects from reference images, which is challenging in real-world scenarios due to the complex geometric features that are highly coupled with the content. To achieve this, we present a generative decoupling method that accurately separates image content and lighting effects. Using two fine-tuned diffusion models, we generate a new million-scale image-content-lighting triplet dataset. We then train the model using IC-Light as a generative model, injecting the triplet with the reference lighting image as an additional conditioning signal. TransLight enables customization and natural transfer of various lighting effects. By thoroughly separating the lighting effects from the reference image, it offers highly flexible lighting control. Experimental results demonstrate that TransLight is the first method to successfully transfer lighting effects between different images, offering more customized lighting control than existing techniques and opening up new directions for lighting harmonization and editing research.
