Abstract: A graphics processing unit (GPU) based cone-beam CT iteration algorithm is proposed. This algorithm takes 3D texture as the discrete model of reconstructed object. It includes several base modules: the forward-projection based on ray casting, the backward-projection by backward mapping projections to each slice of object 3D texture, and the image correction and projection correction by multi-texture blending. Compared with classical TMA-SART algorithm, the proposed algorithm has a faster reconstruction speed in full floating point precision, requires less video memory, and is easier to incorporate prior knowledge and constraints into its framework. These advantages of the algorithm are verified by the experiment of the 3D image iterative reconstruction of Shepp-Logan phantom.