Image transfer of Venus using a fire texture.
To do texture synthesis, we take patches from the original texture, quilt them onto our new image with some overlap, and chose a patch that minimizes the sum of squared differences in the overlap between the patch and its neighbors. To make the process more stochastic, we choose randomly from the N best patches that minimize the difference. We do a min-cut (seam carve) between the new patch and its neighbors to minimize artifacts. To do transfer, we use the same algorithm, but in our score metric we also seek to minimize the differences between the source image and our patches.
In image synthesis, given a texture that was relatively consistent and we had a large number of samples to choose from for each patch, the algorithm yielded realistic results. Artifacts can occur when our texture is too varied, in the case of the grand canyon. Given a texture that was sufficiently varied and a large number of samples to choose from for each new patch, our image transfer algorithm could re-render an image in a way that resembled the original image. It is hard for our algorithm to capture high frequencies and details, simply because our textures do not contain the same details as our original image.
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() |
![]() ![]() ![]() ![]() ![]() |