first test of my Sigma 105mm F2.8 DG DN Macro Art

[lothman]

just unpacked the lens and did some first test shots on a Sony A7r4 f2.8 - I am quite happy
run a stack at 50µm -> Zerene pmax, topaz denoise



crop bottom left


crop center


crop top right


Link to fullsize image

Link to a picture of this waver done by the Mitutoyo 7,5x

[lothman]

and this is a single frame at f2.8 out of the camera on another wafer.

Link to single frame

[chris_ma]

thanks for the sample images.
quite a bit of distortion - is that around 1:1?

[lothman]

thanks for the sample images.
quite a bit of distortion - is that around 1:1?

yes 1:1

[Lou Jost]

Chris, I think Sigma and other lens makers are moving more and more to allow distortion to be undercorrected, which gives them more freedom to correct CA. They figure distortion is easy to correct automatically in processing. Sigma was already doing this with their 70mm Art macro lens. I must say the software does do a good job of correcting it.

[chris_ma]

Hi Lou,

yes, that makes sense, and for most purposes it's an excellent trade off.
In some of my technical uses distortion (or rather the lack there off) is critical (like 1 pixel at the edges), so I'm a bit over sensitive.

I'm planning to rent the sigma eventually and compare it to some reference line scanner lenses, would be interesting anyhow and maybe I can get it to work with good lens profiles.

btw, does anybody know if lens profiles are adjusted for distance/magnification or does one have to create a separate one for each magnification?

[Lou Jost]

Chris, wow, 1px is quite a strict tolerance! The only lenses that I know of that can achieve that are photolithography lenses and some of the special Nikkors like the Repro-Nikkor. What do you use?

[chris_ma]

yeah, it's not strictly necessary, and I probably could use lens profiles (although I imagine it would be hard to create them to the same precision), but it is a good feeling to know that even the RAW files are undistorted and it makes stitching easier.

1 pixel on a 40MP camera is about 0.00025 (or 0.025%) on the edges and 0.0002 on the corners if my math is right.

quite a few line scanner lenses can do this on a FF sensor, for example most of the Mejiro OHT, most of the the Xenon Sapphires, and I guess some of the Rayfact lenses (even though the lack of graphs make it hard to verify).

at 1:1 there's quite a few - seems to be a lot easier there (maybe because of symmetrical designs?).
even the macro varon should qualify at 1:1 from what I can figure.

anyway, back to the topic at hand:
the Sigma 105 sure looks like a winner considering the price and ease of use.

anybody knows if there is a lens profile for the 105mm already?

[ray_parkhurst]

Don't those distortion corrections noticeably degrade sharpness due to interpolation effects? I would think they are similar to a small change in scaling, though the effect isn't global.

[Lou Jost]

Ray, you must be right about that. But the error could be of the same order of magnitude as the error between adjacent pixels caused by Bayer interpolation.

[ray_parkhurst]

Ray, you must be right about that. But the error could be of the same order of magnitude as the error between adjacent pixels caused by Bayer interpolation.

Perhaps same order, but it's an additional degradation, and in my experience it's very noticeable. It is of course additive to the demosaicing, which you can't really do much about.

[Lou Jost]

Just as a side note, in general, random errors are less than additive, so things aren't quite as bad as we might expect.


If noise is random, the total noise is the square root of the sum of the squares of component noise levels.

Noise is the standard deviation. Variance is the square of the standard deviation, and variance is additive. So you add the variances and then take the square root.

So in practice, if there is a random error of about 1 pixel (on the average) due to some cause, and another independent random error of about 1 pixel (on the average ) due to some other cause, the total average error would not be 2 pixels but 1.4 pixels. That's because it is very unlikely that both sources of error would be maximally bad at the same time.

[rjlittlefield]

If noise is random...

True, but I think that Ray is referring to effects that are very far from random.

The problem with resampling to adjust geometry is that there are unavoidably large areas of the image where the shift is around 1/2 pixel. In such areas, unlucky subjects will get softened a lot. The canonical worst case is where pixels start out as a sequence of alternating values, say 0,100,0,100,0,100, which after 1/2 pixel shift must end up as just the average: 50,50,50,50,50,50. Of course most cases are not nearly that bad, but there's always a hit.

So, if you care simultaneously about exact geometry and pixel-peeping sharpness, I think there's no substitute for getting it correct in the optics.

--Rik

[Lou Jost]

I think there's no substitute for getting it correct in the optics.

Absolutely, but my point is that the correection may not be as bad as expected.

I think that Ray is referring to effects that are very far from random.

For very special arrangements of the subject, I agree, but I think the effects should indeed be close to random at the pixel level for typical subjects.

[lothman]

I don't bother about distortion but I like the excellent CA control and the superb sharpness over the whole frame, a good piece better than my old Sony FE90.