The traditional 3D deformable model is inefficient in face reconstruction. To address this problem, a linear deformable model is proposed based on facial landmarks. First, a 2D template-based alignment algorithm isdeveloped to process the correspondence between faces automatically, and a facial linear model is built. Then, adynamic component-based deforming model is proposed to select the most correlative components as the basicspace. Finally, the facial shape is reconstructed by a double deformation framework, a combination of the globaldeformation and local modification. Experimental results show that this method outperforms the conventionalmethods in modeling precision, and the 3D faces generated from real-world photos are rather realistic based on afew landmarks.