StableNormal is a method that reduces the stochasticity of diffusion models to achieve stable and sharp surface normal estimation from monocular inputs. Unlike prior diffusion-based approaches, StableNormal focuses on enhancing estimation stability by reducing the inherent randomness of diffusion models, leading to "Stable-and-Sharp" normal estimates that outperform existing methods in various real-world applications. The method employs a coarse-to-fine strategy, starting with a one-step normal estimator (YOSO) to generate an initial normal guess, followed by a semantic-guided refinement process (SG-DRN) to improve the normals. The effectiveness of StableNormal is demonstrated through competitive performance on standard datasets and various downstream tasks, showing that it retains both stability and sharpness for accurate normal estimation. The method addresses the challenge of high-quality surface normal estimation from monocular inputs by repurposing diffusion priors, which have been recently used to estimate geometric or intrinsic cues. However, previous attempts struggled with stochastic inference and costly ensembling steps. StableNormal mitigates this by reducing inference variance, producing stable and sharp normal estimates without additional ensembling. It works robustly under challenging imaging conditions and is also robust against transparent and reflective surfaces. The method is evaluated on multiple datasets and shows superior performance in terms of accuracy and stability. The main contributions include identifying the conflict between the stochastic diffusion process and the deterministic requirement for geometric cues estimation, proposing a simple yet effective solution, and conducting extensive experiments to evaluate the method's accuracy. The results demonstrate that StableNormal achieves a balance between stability and sharpness, making it a promising approach for surface normal estimation.StableNormal is a method that reduces the stochasticity of diffusion models to achieve stable and sharp surface normal estimation from monocular inputs. Unlike prior diffusion-based approaches, StableNormal focuses on enhancing estimation stability by reducing the inherent randomness of diffusion models, leading to "Stable-and-Sharp" normal estimates that outperform existing methods in various real-world applications. The method employs a coarse-to-fine strategy, starting with a one-step normal estimator (YOSO) to generate an initial normal guess, followed by a semantic-guided refinement process (SG-DRN) to improve the normals. The effectiveness of StableNormal is demonstrated through competitive performance on standard datasets and various downstream tasks, showing that it retains both stability and sharpness for accurate normal estimation. The method addresses the challenge of high-quality surface normal estimation from monocular inputs by repurposing diffusion priors, which have been recently used to estimate geometric or intrinsic cues. However, previous attempts struggled with stochastic inference and costly ensembling steps. StableNormal mitigates this by reducing inference variance, producing stable and sharp normal estimates without additional ensembling. It works robustly under challenging imaging conditions and is also robust against transparent and reflective surfaces. The method is evaluated on multiple datasets and shows superior performance in terms of accuracy and stability. The main contributions include identifying the conflict between the stochastic diffusion process and the deterministic requirement for geometric cues estimation, proposing a simple yet effective solution, and conducting extensive experiments to evaluate the method's accuracy. The results demonstrate that StableNormal achieves a balance between stability and sharpness, making it a promising approach for surface normal estimation.