Partial differential equation (PDE) is one of the main methods for image processing and its significance is usually shown by the corresponding variable model, according to which the PDE can be optimized further to reach ideal results. Based on the classic Beltrami flow for image processing, a new metric tensor model on the image manifold is proposed for image denoising. The Beltrami flow with this metric tensor has clear geometrical significance, which induces the optimal selection method for parameters in the metric tensor. Meanwhile, this model provides a unified framework for the classic PDE based image denoising methods and the optimal selection method for its parameters makes the Beltrami flow have a better balance between smoothing the noise and preserving the edges. The experiment results show that the image denoising quality is greatly improved, especially for the images with abundant edges.