DiffusionLight is a technique that utilizes a diffusion model to estimate lighting effects from a single input image. It employs a pre-trained Stable Diffusion XL model to draw a reflective sphere, then expands it to create a panoramic lighting map. This technology addresses the issue that existing neural network-based methods rely on a limited HDR panorama dataset, resulting in poor performance in realistic complex scenes. The key innovation lies in the discovery of the relationship between the diffusion noise map and the quality of the reflective sphere generation, iteratively generating high-quality reflective spheres; and through LoRA training, enabling LDR models to output HDR format. The technology can produce realistic lighting estimations, especially suitable for outdoor scenes.