In this paper, a novel exemplar-based completion model which is used to remove objects from natural image and texture image is proposed. Firstly the model selects a dynamic size of exemplar based on the local texture information. Secondly, the filling order of exemplar is determined by a cross-isophote diffusion data item. The data item is the result of partial differential equation (PDE) which is derived based on the geometrical property. It considers the extent of edge, so it has a good linear structure preserving property. Thirdly, when processing the textured image, both color and gradient information is incorporated into the similarity function, and then the right texture patterns are preserved. Lastly, a gradient constrained total variation (TV) interpolation is used to reduce seams between exemplars in the completed image. Both theoretical analysis and experiments are given to demonstrate the performance of our model.