Enlarging

[slipe]

I cropped a tiny portion of a photo and then blew it up over 30 times in area with both bicubic smoother and Genuine Fractals. Genuine fractals does a better job. I might try the method in Luminous Landscape and compare that when I get time.

Bicubic smoother:


Genuine Fractals:

[Nagasaki]

An interesting test. I wouldn't say that it was as conclusive as you think. Obviously you prefer the output from genuine fractals. For me neither is great and while genuine fractals seems to pull out a little more detail in places in others it seems to introduce digital artifacts rather than detail.

Viewing from close up and from across the room I don't see either as a clear winner.

Ken.

[slipe]

For me neither is great and while genuine fractals seems to pull out a little more detail in places in others it seems to introduce digital artifacts rather than detail.

There isn't any magic. The problem with blowing up an image is lack of information in the original sufficient for the larger size, not degradation in the upsizing. The rough edges that bicubic smoother improves are more a product of aliasing caused by the original pixel grid. Anything that makes the blown up image appear to have more detail is from artifacts introduced. I would guess some of those artifacts are from anti-aliasing with bicubic smoother.

If you blow any of them up until you see pixels you realize there isn't really any more information or detail. Say you blow up the car in the background to try to read the license plate. If a single bicubic or Lanczos upsample doesn't make the plate readable, GF or SI won't make it readable and bicubic smoother will probably lower your chances of reading it. But zoom back out to a more normal view of the scene and the SI or GF image will give the impression you could read the plate better if you blew it up.

But the artifacts introduced by SI and GF are more subtle and effective than any sharpening. And those artifacts can be effectively sharpened, so the effect maintains its advantage over simple sharpening. Another advantage of the artifacts is that they don't seem to actually degrade the image when really blown up like you can see with sharpening.

Here is one I did about 4 years ago to compare bicubic with SI. It is a crop of about 5% of a 5Mp image. 2% increments work a tad better, but I didn't have a lot of RAM back then and 2% was like watching the grass grow once it started using virtual memory. It is a more realistic blow-up as it relates to printing than the first one I posted. The SI enlargement looks sharper in every way until you really blow it up.



This is a little more realistic comparison of GF and bicubic smoother. I oversized them a little over 20% and then reduced them with bicubic sharper as suggested by the tutorial. That didn't seem to change their relative quality. GF doesn't have the advantage over bicubic smoother that SI has over bicubic, but I think it is a little better. I always thought SI was a little better than GF.

[Nagasaki]

My point about more detail was specifically along the arm. I think I can see hairs on the arm in the GF image these are not present in the BS image. If this detail was in the original image then BS has lost it during the upsizing. If it wasn't in the original then I'm seeing things or GF has introduced a very odd artifact.

However the effect on GF on the rest of the image appears almost like oversharpening and I prefer the slightly out of focus appearance of the BS enlargement.

As you rightly say the informaton just isn't there in the original file so it is purely a question of how cleverly the software fills the gaps when enlarging the image.

I suspect that if you want a poster sized print for your wall either will do the job as long as you don't get too close.

I often project pictures using a 800x600 digital projector. This looks fine from normal viewing distances. If you walk up to the screen you'll see pixels like dinner plates.

Ken

[slipe]

The hairs are there in the original. Bicubic smoother evidently gives up detail to anti-alias the image. You get a smoother looking image but with a loss of detail. Since loss of detail is a problem with a large upsample, I would prefer working with the GF image if I wanted a large print. You can drop the contrast a tad in curves, but you can't get the detail back. There aren't sharpening halos like you get with an oversharpened image, so the GF image is workable.

That great a blow-up was probably a bad idea. The image looks sharper with GF when you view it at a more reasonable size, and it doesn't look oversharpened. I would certainly be happier with a print of the osprey from GF, and would probably be happier yet with SI.

Bicubic smoother probably isn't going to be my standard upsample method.

[Nagasaki]

Yup I'd give the Osprey to GF. Likewise the Si image although on this one it also seems to have preserved thenoise in the background, no such thing as a free lunch.