Natural looking output images are computed from input images based on given user constraints. Pixels in the output images are assigned a shift such that the respective output pixel value is derived from the value of the input pixel whose location is related to that of the output pixel by the shift, at least one shift being non-zero. The shift is determined by an optimization process adapted to minimize a cost function that includes a data term on the shifts of single pixels and a smoothness term on the shifts of pixel pairs. The output image is computed by applying the optimized shift-map between the input and output pixels. The data term can include shift constraints that limit the location in the output images of selected input pixels, and saliency constraints, indicating a preference that selected pixels in the input images will or will not appear in the output image.