StableDelight is an advanced model focused on removing specular reflections from textured surfaces. It builds upon the success of StableNormal, which aims to enhance the stability of monocular normal estimation. StableDelight applies this concept to tackle the challenging task of reflection removal. The training data includes datasets from Hypersim, Lumos, and various specular highlight removal datasets from TSHRNet. Additionally, we integrated multi-scale SSIM loss and random conditional scaling techniques during the diffusion training process to enhance the clarity of single-step diffusion predictions.