This paper addresses the issue of manifold deviation in training-free conditional diffusion models, which occurs when loss guidance is used during the sampling process. The authors theoretically establish a lower bound for the estimation error of the loss guidance, revealing that this deviation can lead to compromised sample quality and longer sampling times. To mitigate this problem, they propose Diffusion with Spherical Gaussian Constraint (DSG), a method inspired by the concentration phenomenon in high-dimensional Gaussian distributions. DSG constrains the guidance step within the intermediate data manifold, enabling larger guidance steps and improving sample quality. The authors provide a closed-form solution for DSG denoising, making it easy to integrate into existing training-free conditional diffusion models with minimal computational overhead. Experimental results across various tasks, including image inpainting, super-resolution, and style guidance, demonstrate the effectiveness and adaptability of DSG in terms of both sample quality and time efficiency.This paper addresses the issue of manifold deviation in training-free conditional diffusion models, which occurs when loss guidance is used during the sampling process. The authors theoretically establish a lower bound for the estimation error of the loss guidance, revealing that this deviation can lead to compromised sample quality and longer sampling times. To mitigate this problem, they propose Diffusion with Spherical Gaussian Constraint (DSG), a method inspired by the concentration phenomenon in high-dimensional Gaussian distributions. DSG constrains the guidance step within the intermediate data manifold, enabling larger guidance steps and improving sample quality. The authors provide a closed-form solution for DSG denoising, making it easy to integrate into existing training-free conditional diffusion models with minimal computational overhead. Experimental results across various tasks, including image inpainting, super-resolution, and style guidance, demonstrate the effectiveness and adaptability of DSG in terms of both sample quality and time efficiency.