Re: I added bicubic examples. (nt)

Ah, I think I know what part of the problem is. Your sample image was
calculated by simulating the behaviour of a camera with finite-area
sensels, but no anti-aliasing filter and with an infinitely sharp lens.
So you have a number of aliasing artifacts in the image.

If you look at the spokes of the bicycle wheel in the small unsharpened
sample image, some of them look like a string of beads with gaps between
the beads, rather than the continuous piece of metal that they are.
When you upsample this using bicubic interpolation, you get "jagged"
edges on the spokes - because that faithfully reproduces the jagged
appearance of those spokes in the small image.

If the test image was properly filtered before sampling, the spokes
would look less sharp, but they should appear continuous. And if the
spokes look continuous in the source, bicubic interpolation will
preserve that appearance, without adding ringing around contrasty
edges.

Even with the limitations of the test image, the bicubic upsampled
versions (both of them) look much better to me than the ones done with
data-dependent Lanczos (both of them). The bicubic ones are merely
somewhat fuzzy, but faithfully reproduce the size and appearance and
position of objects in the original. The DDL method creates patterns
where there were none (the clown's shirt, the road texture) and distorts
shapes and edges (look at the trees, or buildings in the background).

Dave
.