Enlarging images for printing

[ChicagoJohn]

I apology in advance if this topic has been thoroughly discussed previously and/or if this is not the right forum. But I have been reading a bit about this subject lately on the internets.

The problem would be framed as how to change the native pixels per inch in an image that is of technically high IQ to a level that will produce a suitable result in larger print size than could be achieved at, say, 240 pixels per inch from the native resolution.

For example, if one wanted to print a 6MP image of 3000 X 2000 at 240 ppi, that would equate to an 8 X 12 inches / 20 X 30 cm. If the size is changed to 20 X 30 inches / 51 X 76 cm, without resampling, the pixel density drops from 240 to 100.

So my question is what actual experience folks may have and recommendations to make as to how to optimize enlargement. So far, I've found comments like;

1) depends on the medium -- rough or glossy -- and the distance from which it'll be viewed

2) use PS CS5 Resample with Bicubic Smoother in increments of 10% to get to the desired enlargement

3) use PS CS5 Resample with Bicubic Smoother to a size larger than needed then resize down to the target with Bicubic Sharper and then use a sharpen filter to further remove the resulting softening

4) Buy a program / plug-in specifically for this purpose.

I'm trying to take images to 20 X 30 inch size using professional printer such as WalMart or Costco. So if anyone has experience with this and has had good results, I'd like to hear about how.

[blumesan]

I often use Costco for prints, large and small. My routine procedure is to upsample the image (when needed) to 300ppi using bicubic smoother, one step. Then while viewing the image at 100% apply output sharpening (smart sharpen) until it just about appears a bit oversharpened. I always specify "no color correction" since I have already done the work by softproofing using Costco's printer profile.

If I am using a higher end lab to make the print, I believe their software is usually better than PS to accomplish the needed change in resolution. I just send the native file (sharpened as above) and specify the required print size.

[ChicagoJohn]

Originally Posted by blumesan

I often use Costco for prints, large and small. My routine procedure is to upsample the image (when needed) to 300ppi using bicubic smoother, one step. Then while viewing the image at 100% apply output sharpening (smart sharpen) until it just about appears a bit oversharpened. I always specify "no color correction" since I have already done the work by softproofing using Costco's printer profile.

If I am using a higher end lab to make the print, I believe their software is usually better than PS to accomplish the needed change in resolution. I just send the native file (sharpened as above) and specify the required print size.

Hi blumesan,

I use Costco too and follow the "no color correction" option. Before reading your post this morning, last night I tried the multiple 110% enlargement technique with Bicubic Smoother, taking an imaged I wanted to print at 20 X 30 from 16 MP up just past 9000 X 6000 pixels (8 iterations) and then back down to that target using Bicubic Sharpener. It looked reasonable so I set it in for printing and will report the results later this evening.

Anyway, after reading your post, I decided to do a side-by-side comparison of the single stage and multiple iteration methods. The image I used was one I took April 8 and posted a few days ago on the Cityscape and Landscape forum here.


Edge of The Bean- Millennium Park Chicago by ChicagoJohn, on Flickr

I did a radical crop on the images produced by each method from the top edge of the frame for the building on the extreme left. The top one is the single stage enlargement and the bottom one the multi-stage, 110% method.


comparison of enlargement iterations by ChicagoJohn, on Flickr

I can see a marginal, but I think significant difference in that there is less aliasing with the multiple stage method. I seriously doubt that this would be visible at the distances a 20 X 30 is ordinarily observed, and perhaps only under magnification of the print, but it does appear to be there.

Also, thanks for the idea of using a higher end lab. I'll PM you to get some recommendations. I think I'll call Costco as well to see if they have any software they use when a lower native resolution file is submitted for a much larger print.

POSTSCRIPT

After examining the Costco 20 X 30 in print and putting it in its frame, I couldn't be happier. If one gets to the 4-5 inch position in front of it, one can begin to see some softening and artifacts, if, that is, one knows what one is looking for. Looking at it under a 5X jeweler's loop, one sees the results shown above under radical magnification, but it is of no practical significance.

I did check with Costco when I picked up the print, by the way, and they have no process that intervenes between native resolution and a requested enlargement. So maybe higher end labs would be able to do that, and I will look into the cost-benefit equation there.

Oh, I almost forgot -- this is the first time the processing lady at Costco told me, unsolicited, "Awesome picture". That made me pretty happy to hear since she does this all the time.

[ChicagoJohn]

This morning I downloaded the 30 day trial professional version of popular, patented plug-in enlargement software said to be based upon fractals and performed an experiment on the same file used for the comparison previously reported.

Comparing the results from this plug-in (which took 10 minutes to run on my computer) with the results from the 110% iterative method (which took under 3 minutes) at 400%, just at the borrderline of pixelation, examining several areas of the image, I can see no difference in the two results except that in a few areas the plug-in shows very slightly more pixelation. However, with respect to resolution of detail and degree of artifacts, they are virtually identical.

Of course, you can do batch processing with the plug-in, but I may need to do this maybe once a month, if that, and it should be easy to automate the 110% iterative method as an Action in photoshop.

[apaflo]

ImageMagick is a free software package that runs on every major platform, and performs virtually any command line image editing function one can imagine. That makes it very nice for batch processing of course; and depending on a GUI editor's scripting ability it may be possible to invoke some "batch" processes on single images from an editor too.

The significance for this thread is the diversity of filters available for resizing images. There are currently 26 different filters available. Plus there are adjustments possible for most filters (radius, blur factor, lobe count, etc).

Essentially, to whatever extend you wish to study the technical aspects of resampling, ImageMagick tools will allow adjustment and experimentation.

In addition to the filters and their configuration, other significant parameters can be manipulated too. For example, not all image editors have resampling algorithms that allow changing the bit depth or the gamma correction before resampling and then resetting afterwards.

I use a script, which invokes the "convert" tool from ImageMagick, to resize images. This is my default for an image that is down sized (in this case to a width of 1000 pixels):

convert input_image.jpg -depth 32 -gamma 0.454545 \
-filter Lanczos -sampling-factor 1x1 -resize 1000 \
-quality 90 -gamma 2.2 -depth 8 -density 360 \
-units PixelsPerInch output_image.jpg

Note that the default for downsizing is a Lanczos filter (upsizing defaults to a Mitchell filter), and the bit depth is first converted to 32 bits for precision while gamma correction is removed to allow resampling as a linear data set and then restored afterwards, as is the 8 bit-depth. I set DPI to 360 because I use Epson printers. Also, with some conversions (typically going from PostScript or PDF to TIFF formats) it can be important to set "-type truecolor" too.

This provides what appears to be unmatched flexibility and quality for image resampling.

[ChicagoJohn]

Originally Posted by apaflo

ImageMagick is a free software package that runs on every major platform, and performs virtually any command line image editing function one can imagine. ...
This provides what appears to be unmatched flexibility and quality for image resampling.

Thanks so much, apaflo! I'll download that today and check it out. Good to know there is a lot more to explore here.

Would you be willing to post some examples which compare the results of, for example, a single-step bicubic smoother in PS, the 110% method, perhaps the commercial software I mentioned, against the results you achieve using the ImageMagick software? That would be helpful empirically if you don't mind doing it. Otherwise, I will give it a try.

In addition, a technical question arises in my mind from reading your post. You mention conversion to 32 bit depth for improved precision. Bit depth, in my understanding, involves the gradations of luminance recorded in a sensor. That would not then have any bearing upon sharpness, artifacts, etc. in the interpolation process, would it? Also, if one starts out with 8 bit depth, then converting it to 32 bit depth would not seem to provide any additional information regarding luminance.

I'm new to this and would appreciate anything you would care to share or information you could point me to.

[apaflo]

Originally Posted by ChicagoJohn

Thanks so much, apaflo! I'll download that today and check it out. Good to know there is a lot more to explore here.

Would you be willing to post some examples which compare the results of, for example, a single-step bicubic smoother in PS, the 110% method, perhaps the commercial software I mentioned, against the results you achieve using the ImageMagick software? That would be helpful empirically if you don't mind doing it. Otherwise, I will give it a try.

Unfortunately I'm not in a position to accomplish that in any meaningful way. For one, the place where it actually makes a difference is when making large prints, not when downsizing for the Internet. Another problem is that results vary from one image to another, so one image might be significantly affected while the particular one you try might not be.

Here is a really really interesting article that describes why images should be converted to a non-gamma corrected linear data set before resizing. They provide downloadable images that show whether a given editor is or not doing that.

Gamma error in picture scaling

This article is a good discussion on such things as the USM and Sharpen tools used by ImageMagick and compares them to GIMP and PhotoShop.

Sharpening using Image Magick Red Skies at Night

In addition, a technical question arises in my mind from reading your post. You mention conversion to 32 bit depth for improved precision. Bit depth, in my understanding, involves the gradations of luminance recorded in a sensor. That would not then have any bearing upon sharpness, artifacts, etc. in the interpolation process, would it? Also, if one starts out with 8 bit depth, then converting it to 32 bit depth would not seem to provide any additional information regarding luminance.

That is true as long as nothing else is done.

But of course given all the calculations involved in changing the gamma correction and interpolating upsized data points, both speed and precision are affected, though probably not by enough to be obvious without contriving something designed to show it. Generally 16 bits is recommended, and is probably great enough to avoid visible artifacts, but since the computer works with 32 bit data I chose to just let it work with the native data size. If it was doing I/O to disk that would be slower, but this is all in memory and should be faster.

I'm new to this and would appreciate anything you would care to share or information you could point me to.

ImageMagick also has an article on how to resize images that is pretty interesting too:

Resize or Scaling -- IM v6 Examples

[sbingham]

Been here many times! I print 24" x 36" and 16" x 24" on my Epson printers.

1- Using PS CS5 and Epson printers set Image/Image Size to 240 pixels per inch (no need for 300)

2- Make sure all three boxes are checked. (Scale Styles, Constrain Proportions, Resample Image)

3- Select Bicubic Smoother

4- Choose a new Document Size (in inches or metric). This typically takes less than ONE second on my computer to convert.

5- Print

After extensive testing over many years I have found 3rd party software never has equaled Photoshop resampling. There is also absolutely NO need to do this 110% at a time. In fact, this sometimes can be determental. This is pretty much an urban legent dating back to very early versions of Photoshop. Legends die hard!

Although I am now retired I taught photography and Photoshop for 33 years (high school and then college). Now I just shoot for FUN. No more commercial stuff.

[apaflo]

Originally Posted by sbingham

Been here many times! I print 24" x 36" and 16" x 24" on my Epson printers.

1- Using PS CS5 and Epson printers set Image/Image Size to 240 pixels per inch (no need for 300)

2- Make sure all three boxes are checked. (Scale Styles, Constrain Proportions, Resample Image)

3- Select Bicubic Smoother

4- Choose a new Document Size (in inches or metric). This typically takes less than ONE second on my computer to convert.

5- Print

That is a technique which works, but is far from optimum because it ignores one very simple essential consideration: The Epson printer will in effect resample your image from 240 PPI to 360 PPI every time. That will result in a difference between what you can preview on screen (for example sharpening) and what the final print will look like.

For example, if you have an image that starts life at 24.5 MP from a Nikon D3X or worse yet a 36MP image from a D800, The pixel dimensions of the D3X are 6048x4032, and an Epson printer will not resample it if printed at 16.8x11.2 inches. The pixel size for the D800 is 7360x4912, and the unresampled print size is 20.4x13.6 inches. Lets consider what happens when prints are made at the two cited sizes, 24"x16" and 36"x24".

A 240 PPI image for 36x24 will be 8640x5760. That is the size at which it will be previewed, and sharpening applied. But when printed on an Epson at 360 PPI it will first be resampled to 12960x8640. Unless the amount of sharpen applied at the smaller size appeared to be excessive, the print will appear less sharp than it could be. But the difference is small and certainly would not be noticeable without doing prints both ways and making side by side comparisons.

A 240 PPI image for 24x16 will be 5760x3840. Note for both a D3X and a D800 that will require down sampling because the out of the camera is larger. Then of course the Epson printer will resample it again to a larger size, 8640x5760. Not only will the amount of sharpen applied be off, the initial downsizing reduces the real resolution which is then "restored" by upsizing convolution. (The whole purpose of buying high resolution cameras like the D3X and D800 is to avoid exactly that!)

The moral of the story is that resizing an image to anything other than the "native resolution" of the printer is fraught with these hidden perils!

There is one small exception to that, which has to be mentioned but should also be ignored. That is that many current model printers can also produce "high resolution" images. For Canon and HP printers the normal PPI is 300 but they can also print at 600 PPI, while for Epson normal is 360 PPI and high resolution is 720 PPI. First, the higher resolution has little if any effect for photographs, as opposed to line art drawing or text, and second because of the way the printer resamples entire lines rather than per pixel it makes virtually no difference if the actual pixel dimension of the image you send to the printer is at normal resolution or twice that.

Ultimately, for Canon or HP printers image should be resampled to 300 PPI and sharpened by observation at that size, while for Epson printers the image should be resampled to 360 PPI and sharpened at that size.

After extensive testing over many years I have found 3rd party software never has equaled Photoshop resampling. There is also absolutely NO need to do this 110% at a time. In fact, this sometimes can be determental. This is pretty much an urban legent dating back to very early versions of Photoshop. Legends die hard!

That is very true. The reason it appeared to produce better results years ago was only due to the lower level of sophistication available in resampling algorithms compared to today. In fact today, perhaps not with Adobe products, but certainly with the tools from the ImageMagick package, there are resampling algorithms that are just as good (as in, they are the same) as those use in the very expensive RIP's available for commercial printers. And they are vastly better than allowing any print driver software to resample with what amounts to something very similar to a "nearest neighbor" method.

[sbingham]

You know you might very well be right. Granted the native resolution of Epson printers is either 360 0r 720 depending on the printer, however I have found through emperical print testing very little difference when using an even fraction such as 240. Eric Chan, who now works for Adobe as the number 1 guy behind Thomas Knoll in ACR, has pretty much stated the same thing. He is regarded by many as THE Epson printer expert. I also realize that Epson must resample to its native resolution before printing. I open all my Nikon raw files using ACR set at 240, whatever my camera is at the time.Using USM I usually view at either print size, 50%, or 100% (rarely). My final sharpening is always done for print size so I am not sure your theory would apply in my case.