A new method using Gaussian Scales Mixtures in the Curvelet domain is proposed for image denoising. The Curvelet local coefficient model is first built using Gaussian Scale Mixtures. Then, the coefficients are estimated using Bayes least squares estimator. This algorithm combines the merits of Curvelet for image edge representation and Gaussian Scale Mixtures for capturing correlation of local coefficients. Some numerical experiments show the effectiveness of our method for image donoising, especially, for texture and detail predominated images.