The algorithm comes from the Wei-Levoy method but forces copy of entire parts of the texture sample. To do this, the comparison of the current neighborhood is no more done with all the input neighborhoods (or with the nodes of a tree representing all the input neighborhoods) but with a small number of neighborhoods chosen as follows:
(a2) (b2) (c)
Ashikhmin synthesis. The current pixel is marked by in the output image (a2). The pixels in the current neighborhood were assigned with the values of the pixels in the input image (c) (see the gray arrows). The pixels candidates in the input image (a1) are marked by and (b1) represents then with their respective neighborhoods. Those neighborhoods are compared with the current one (b2), and the pixel with the most similar one gives the value to assign in (c).
As for the Wei-Levoy method, it is possible to introduce a multi-resolution method, using the same principle.
See the results of this method.