While not directly about the camera, this is one in a series of posts that relates to the Nikon Z6. You should be able to find all the posts about that camera in the Category List on the right sidebar, below the Articles widget. There’s a drop-down menu there that you can use to get to all the posts in this series; just look for “Nikon Z6/7”. This is also relevant to the Fuji GFX 100; for other posts about that camera, look at Category “GFX 100”. This is also a continuation of testing that I’ve been doing on the Epson P800 printer. I’ve created a category called “Printers”, and put this post in that category. If you go to the Category List (on the right in the desktop formatting), find “Printers” and click on it, you’ll see all the posts in that series.
In the previous five posts, we looked at the results of resizing files from three different camera/lens combinations:
- Fujifilm GFX 100, Fuji 110 mm f/2 lens, set to f/5.6. That’s not the sharpest aperture for that lens, but the sharpness is not far down from its best there. Lightroom sharpening amount = 20, radius = 1, detail = 0. Image height is 8776 pixels and image area is 101 megapixels.
- Nikon Z6, FX mode, Zeiss Otus 85 mm f/1.4 lens, set to f/4. That’s not the sharpest aperture for that lens, but the sharpness is not far down from its best there. Lightroom sharpening amount = 30, radius = 1, detail = 0. The increase sharpening is to compensate for the (weak) AA filter on the Z6. Image height is 4000 pixels and image area is 24 megapixels.
- Nikon Z6, DX mode, Zeiss Otus 55 mm f/1.4 lens, set to f/2.8. That’s pretty close to the sharpest aperture for that lens. Lightroom sharpening amount = 30, radius = 1, detail = 0. The increase sharpening is to compensate for the (weak) AA filter on the Z6. Image height is 2780 pixels and image area is a bit less than 12 megapixels.
I found that Gigapixel AI was the best resizing tool for these images, and that, even using that tool, differences between the two Z6 captures and the GFX 100 one were, for a 1 meter high print, apparent from dome distance away.
A reader who saw the previous posts asked a related question: could a human with ordinary sight see the difference between a 360 pixel per inch (ppi) print and a print of lower resolution. I took the GFX 100 image, resized it to 240 ppi and back to 360 ppi with Photoshop, using “Automatic” for the interpolation method. Then I created another copy, resized it to 180 ppi and back to 360 ppi. I printed all three with an Epson P800 printer on Premium Glossy Photo Paper.
It was trivially easy to tell which was which, even at arm’s length. All you had to do was glance at the Siemens Star. So the human eye’s close focusing distance is not an issue here. As I said earlier, I am no eagle eye. My vision in my left eye is 20/20 most of the time, although sometimes I get to 20/15. My right eye is about 20/25.
My next project was to take the GFX 100 image, resize it to 720 ppi without resampling, resample it to 540 ppi, and save it. Then, starting with the 720 ppi image, resample that to 360 ppi and save it. Then I printed all three images out to the P800 with the 720 ppi (Finest Detail) option checked, letting Lightroom’s Print module do the upsizing for the 360 ppi and 540 ppi images.
What a difference! All three images were very close to each other in quality, even using a loupe.
My second conclusion is this: the P800 (and every other Epson printer I’ve tested for this) is limited to an effective resolution of about 360 ppi, even in 720 ppi mode.
Corollary to the above, when taken together with the other tests in this series: if detail is of paramount importance, it’s best to capture the image with sufficient resolution that you will have at least 360 ppi at the desired print size.
Still waiting on that 112mpx sensor in a Sony 135 format camera to get me that 24×36 in print at 360 dpi. Thanks Jim!
*ppi
Thanks!
The “common wisdom” I’ve heard is that one needs 8×8 dots per pixel. You are saying that 4×8 has clear visible advantages over 6×12. Makes one wonder where the line actually goes ;-).
I would have thought that they would try to detect B&W, and use a different dithering algo there — but you say that this works with the star as well… Still, I expect that a suitable B&W mode may give better results…
[BTW, your boilerplate «in the Category List on the right sidebar, below the Articles widget» does not make a lot of sense when there is NOTHING on the right! 🙁 — But I have no clue how to deal with variable layout in my presentations as well.]
The rule of thumb you quoted is from the days of screening. Using error diffusion with blue noise dithering, It’s more complex than that.
Sorry, now I found
https://blog.kasson.com/technical/more-on-epson-driver-resampling/
— so I’m completely lost. It looks like you were saying that reproducing data at 480ppi was not a big stress on (a certain?) printer. The results were clearly distinguishable both in high contrast and low contrast setting.
Could you help me untangle myself? Please explain what THIS post and THAT one “mean together”! Thanks!
The differences are subtle and subject-related. It is possible to construct tests where anything less than 720 ppi produces artifacts.
Also note that the link you posted referenced tests where the driver did the resampling. In the recent tests I posted, Lr did the resampling. Lr is much better that this than the driver, which uses nearest-neighbor.
Jim, sorry, but this did not help… Let me try to summarize my confusion; here is how I read what you wrote:
• In 2011, you wrote that printing at 720ppi, the data with spacial frequency of 240lp/in is reproduced so that it is discernible (maybe with a loupe) even if at low contrast. (But, obviously, this does not work when printing at 360ppi.)
• Now you write that printing at 360ppi and at 720ppi gives very similar results — even with a loupe, and even if the subject contains a Siemens star.
You even say: “is limited to an effective resolution of about 360 ppi, even in 720 ppi mode.”
How did I get mixed up so much?!
You are still ignoring the resampling differences. The earlier post you linked to was a test that was particularly designed to work with the nearest-neighbor resampling done by the Epson driver.
If you want to see what happens when you resample in Lr, look at the bottom of the following page for some pictures:
https://blog.kasson.com/gfx-50s/mp-resizing-printing-scanned-samples/
You will see that, at high magnification, there are some. You will also see that they are subtle.
Jim, I still think you are trying to INFER something which I do not get.… ;-(
I would really appreciate some DEFINITE answers! Let me try again to summarize how I try to interpret what you wrote:
• There is a workflow which can put high-resolution (say, 240lp/in) details to Epson printers.
• LR cannot.
Is it correct? (Maybe it goes so slow since I would not care less about LR?)
You see, I think I’m missing much more fundamental that what you wanted to address!
Even Lr can put 360 ppm details down on the P800, I demonstrated that in the posts in the current series, in which there is a significant difference between the 240 ppi and the 360 ppi images. If you want to resample less that 720 ppi raster images that come from vector graphics front ends, and wish to retain the most detail at the expense of introducing some alisaning, then nearest-neighbor could be the best interpolation method. That’s the one the driver uses.
If you have vector graphics, it is entirely possible that rasterizing them at 720 ppi will give significantly better results than 360 ppi. You can see in the posts in the latest series that with Lr upsizing and photographs made with a very sharp camera — the GFX 100 — that, while in magnified view 720 ppi is very slightly better than 540 ppi and 360 ppi, the difference is slight. When I look at the prints with the naked eye, I find the differences virtually immaterial.
This is in response to your statement: “My second conclusion is this: the P800 (and every other Epson printer I’ve tested for this) is limited to an effective resolution of about 360 ppi, even in 720 ppi mode.”
This topic has been of interest to me for a long time, significantly before I produced these 2013 graphics:
https://pbase.com/misterpixel/image/153408976
https://pbase.com/misterpixel/image/153408983
https://pbase.com/misterpixel/image/153408980
Please read the text below these images to see what they represent and how and why they were fabricated.
Strangely enough (at least to me) this experiment seems even more valid now in 2019 where high resolution photo sensors with no AA filter, combined with exceptional lenses, are becoming omnipresent.
Admittedly this was printed on an Epson R1900. I now use an Epson P400 but don’t think it would make any difference. Only the color science is dissimilar and not the fundamental underlying resolution capability.
There are two areas of unease on my part. The first is more or less irrelevant and that is the poor quality of the displayed scan at 2400 dpi. The second, and most important, is my logic in insisting that a 720ppi image can be visually discernable or at least that this experiment displays that.
Yeah so where’s our 720 dpi printer?!
DUH…brain cramp on my part:(
That’s what I get for looking back on a technical exercise I did many years ago, not spending much time thinking about it and drawing correlations to Jim’s experiment in real photographs. While it’s true that the Epson printer engines can work at levels significantly over 360 ppi and have the potential to show that under ideal circumstances, I was having a difficult time visually relating Jim’s and my results. It took looking at Jim’s new testing, reviewing my old experiment, looking at old lens testing I had done using artificial charts, and even rasterizing a few new graphics differently to realize what was really going on. There are two important factors here and given my past experience, both of them should have been obvious to me.
First and most important was that I was using a rasterized vector graphic to represent a real world photograph of a hard edge image AND my graphic had been rastered with absolutely NO ALIASING. That’s OK for making my point re: printing but not even non-AA sensors combined with great lenses and looking through significant air can truly approximate this; especially with a Bayer sensor. I also suspect the printer driver (by presumably using some form of ‘nearest neighbor’) helps preserve the delicate contrast in the stair-stepping of a non-aliased graphic when printing.
And secondly… this stair stepping becomes much more obvious when printed and SIGNIFICANTLY ‘enlarged’ like I did when I scanned with my Epson 4990 flatbed. Admittedly I could see this easily in the print with the naked eye because it was not aliased. It’s easy to forget how much difference there is between a complex graphic rasterized with and without aliasing at 360 ppi. While our eyes can pick up on this artificial stair stepping fairly easily when printed well, it is not germane above 360ppi for printing of most (or any) real world photographs as Jim’s scans show. That miniscule sharp and ultra high-contrast systematic stair-stepping essentially doesn’t exist in a photograph.
OK, now I think that I understand what is going on. (I could not do it before, since I did not realize somebody might have wanted to feed a printer at the non-native resolution ;-).) Let me try:
• Epson has a very good potential for high-resolution printing, PROVIDED one feeds it data at exactly 720ppi.
• LR down/upscaling uses a very strong (or naive?) anti-aliasing filter, so printing from LR kills the high-frequency details.
Myself, I use custom pre-filtering when resizing-for-printer (I hand-craft a polynomial for the given ratio; — I can do it since I print rarely), so I did not encounter such problems in my workflow. (It goes through 254ppi printing in El-Co, and 100lp/in details come out “reasonably” well — for some value of “reasonable” ;-).)
(BTW, I do not know why you mention GFX 100 in this context. This was about dpi, not MPix, right? E.g., it still makes sense to ask: “Can one print the photo from GFX 100 on 5×7in so that 240lp/mm details are visible with loupe?”)
WRT the efficacy of printing at 720 ppi with various sampling techniques, have a look at this:
https://blog.kasson.com/technical/printing-at-28801440-dpi/
Also this:
https://blog.kasson.com/technical/resampling-for-printing-%e2%80%93-summary/
Important excerpt: “With conventional error diffusion techniques, the printer should be capable of marginally resolving 1440 line pairs per inch in one direction and 720 line pairs per inch in the other. Used with the Epson driver, it cannot. Downhill with a tailwind, it can resolve 360 line pairs per inch, and it doesn’t do that well. The best dependable resolution is 180 line pairs per inch. How come? Two reasons:
First, the driver resamples the image before it or the printer applies the error diffusion. When the Finest Detail option is checked, the image is resampled to 720 pixels per inch, and otherwise it’s resampled to 360 ppi. The resampling algorithm is nearest neighbor.
Second, there is enough noise (dither) added to the error diffusion algorithm that a single pixel line at 720 ppi occupies about twice that width on the paper.”
I wouldn’t say that the Epson printers have great resolution at 720 ppi. It’s certainly a long way from being able to resolve a 360 line pair/inch target:
https://blog.kasson.com/technical/injet-printing-on-epson-part-3/
I should probably retest with the P800.
I think now (after this explanation, and the recent new samples) I start to see what causes the confusion. Let me try again to summarize:
• Epson driver has a bug: it spreads ink by 1/1440in wider than needed (on all 4 — or 8 — sides);
• In addition to this, it does not compensate for the resulting darkening of the image.
These two effects result in a reduced MTF on high spacial frequencies, AND in the image being too dark in zones with high contrast on high spacial frequencies.
Note that the reduced MTF ALONE would not present much problem (unless in zones of very high contrast)! Doubling of the width of a line results in MTF of cos(f·π/2) — which is easy to compensate up to f about 0.7 (which would be about 250lp/in). Even in the diagonal direction, where it is better described as cos(f₁·π/2)cos(f₂·π/2) it would not be a big deal. (This explains your results from 2011, when you printed 270lp/in without a big-deal-degradation.)
However, darkening is harder to compensate without knowing which tonal curve Epson uses in their conversion for brightness-to-ink-amount… (As usual, this would be “trivial” to fix if the source code was available for modification… — It looks like Epson decided that this “widening” is beneficial in some regards, but instead of diffusion of ink they used duplication of ink…)
I don’t think this is a problem with the driver, but with the ink droplets being too large.
I have no objection against the theory that it is due to ink spreading — but still, the driver should know EVERYTHING about the spreading of ink. I still suspect that the driver COULD have compensated for blackening (by decreasing the amount of ink).
But indeed, with the experiments you did, one cannot exclude the possibility that the driver is doing “the best it can”. The extra test needed is to print the pattern similar to what you did in “Dot gain…”, but with grayscale lines. For example, modulate the blackness of these lines by a diagonal gradient.
It may happen that hairlines may have only very discrete ”blackness”. However, IF the driver can reproduce the gradient, this would mean that it also COULD have compensated for blackening due to ink spreading…
This started out as a discussion about how large you had to print before the resolution of the GFX 100 got to be more useful than the resolution of lesser cameras. My opinion is that above 17 inch high prints, you’ll see a demonstrable advantage with the GFX 100 over the GFX 50S in resolution. Of course, for the GFX’s relative freedom from aliasing, to help, you won’t have to go that large.