Quality issues with Distiller Downsampling
Problem description
It seems that Acrobat Distiller when configured to reduce the resolution of image objects included in PostScript data streams (so called "Downsampling") uses sometimes the "Subsampling" interpolation method even if the joboptions in use should ensure that "Bicubic downsampling" will be used (for a comparison of available distiller downsampling methods see below).
From a quality point of vue only the "Bicubic downsampling" is suitable to reduce the size of images inside PDF files. The silent fallback to the "Subsampling" algorithm results in a massive decrease of image quality.
From our testings it seems that Distiller uses the Subsampling algorithm in cases where the downscale ratio is between 50.x% and 99.99%. If the ratio is exactly 50% or below then the downsampling method configured via the joboptions will be used so the problem normally does not occur in all cases but only in some limited (which makes it hard to reproduce if one doesn't follow exactly the right steps regarding the effective image resolution)
Examples
Original picture (300 ppi resolution, placed in layout application with 74% scaling ratio which leads to an effective resolution of 405 ppi):
Image after Destillation using "PDFX-ready-coatedSheet_Dist7.joboptions" from swiss based pdfx-ready.ch initiative. Due to the DownsampleTreshold setting of 1.33 and a destination resolution of 300 ppi the image in the PDF file has been downscaled from 405 ppi to 300 ppi -- obviously not using the "Bicubic downsampling" algorithm but something producing low end quality – enjoy the horizontal line pattern and other artefacts especially on edges/angles:
In contrast to the image above the same downscaling applied (74% from 405 ppi to 300 ppi) in Photoshop using Bicubic interpolation:
An image with an effective resolution twice as much as the downsampling resolution. Distiller uses the interpolation algorithm specified in the joboptions resulting in no visual problems
How are commonly used Distiller joboptions involved?
As we've seen above the critical downsampling range is within 50.x% to 99.99%. This can happen if the so called Downsample Treshold is below 2.00. The Joboptions (distiller settings) contain different treshold settings for color, grayscale and bilevel images. Default Distiller settings from Adobe set all three values to 1.50 -- see the "High Quality Print.joboptions" for example:
/ColorImageDownsampleThreshold 1.50000
/GrayImageDownsampleThreshold 1.50000
/MonoImageDownsampleThreshold 1.50000
This means when using Adobe default Distiller settings images that have an effective resolution between 150% and 200% compared to the destination resolution happen to be shred by distiller (using silently the subsampling algorithm instead of bicubic interpolation as configured by the joboptions). Or in other words: the problem described above applies to all images with an effective resolution between 450 and 600 ppi (that means an image with 300 ppi native resolution scaled down to 50.x% - 66.6% in a layout application!)
The situation gets even worse when using distiller settings from noted PDF experts or initiatives like the aforementioned pdfx-ready group. Those joboptions set a DownsampleTreshold of just 1.33 resulting in a wider range of images that are affected by Distiller's downscale bug. Images with an effective resolution between 133% and 200% of the downsampling resolution will be shred (note: that applies to images with native 300 ppi when scaled down between 50.x% and 74.x% in a layout application).
Since Distiller 5 end users are able to set these Downsample values within Distiller or to disable downsampling completely. But as a matter of fact especially users with lower prepress knowledge will simply use unchanged default settings provided by Adobe or 3rd party authorities. So those people not knowing that much about prepress-appropriate PDF creation blindly trusting widely spread distiller settings from noted experts happen to provide PDFs to printing bureaus which contain shredded image material not even knowing anything about the dangerous settings they use
Steps to Reproduce
Simply use our Downsample-Testbild.eps which has a native resolution of 405 ppi by either dragging it to Distiller and using a DownsampleTreshold of 1.33 or below. Or experiment with a higher DownsampleTreshold and ensure that the effective resolution of your test images is within the dangerous range (between DownsampleTreshold * downsampling resolution and 2 * downsampling resolution)
Solution
- Adobe should ensure that the downsampling algorithm configured in joboptions will be used in every situation and that no silent fallback to the worst algorithm possible (Subsampling) happens any longer
PDF experts/initiatives/boards – and of course Adobe itself – should take care that joboptions they provide with quality output in mind never ever contain a DownsampleTreshold below 2.0 and that Downsampling is disabled by default.
Default settings should work as expected without requireing any sort of manipulation by an end user prior to using them. That means when the goal is "quality output" then everything that prevents that should be disabled in default settings. If an end user then makes changes to these defaults then it's his/her own decision hopefully based on appropriate knowledge regarding potential consequences. But it should'nt be necessary to fine tune settings prior to productive use as it is the case in most production environments these days (almost everyone with a bit prepress knowledge turns downsampling immediately off. Why should this step be necessary?)
A small footnote to DownsampleTreshold vs. PDF file sizes (a common misunderstanding). If you reduce the DownsampleTreshold value you won't get automatically PDF files smaller in size because it depends on the effective resolution of each and every single image in the PDF file. The difference between DownsampleTreshold 2.00 and 1.33 for example only affects images with an effective resolution between 133% and 200% compared to the configured downsample resolution. All images with a resolution of 133% and below won't be downsampled at all and all images with a resolution of 200% and above will be downsampled in both cases. That means only a pretty small amount of images will be affected by lowering the DownsampleTreshold. And those images are the ones where a size reduction has less effect compared to images with eg. 200% – size reduction to a fourth – or 300% – reduction to a ninth – that will be downsampled in both cases (we made a small comparision between 7000 identical catalog pagues. PDF file size with DownsampleTreshold 1.33 is 38657 mb, with 2.0 just 38790 mb – only a few 133 mb more and the guarantee that no images have been destroyed by a wrong scaling algorithm.
Short summary
Adobe's Distiller uses not the algorithm configured by joboptions when downsampling images to a final resolution between 50.x% and 100% of the effective resolution of the image in question (effective resolution --> native resolution divided through scaling ratio in layout application) but instead the subsampling algorithm which produces visual defects and massively decreased image quality.
Adobe should fix Distiller to honour the joboptions' downsample settings in every situation.
Simply relying on Adobe and a future Distiller update isn't enough and therefore PDF experts/boards/initiatives should ensure that publicly available joboptions they provide do not contain DownsampleTreshold settings below 2.0 and disable Downsampling per default.
Additional Material
Sample files:
Below some files to test with. The TIFF images (RGB PNG versions above) are extracted from a destilled PDF file with the touchup tool except the last one which has been created using Photoshop:
Distiller interpolation algorithms:
Distiller knows three different interpolation algorithms. For quality purpuses only "Bicubic downsampling" should be used. Quoted from Adobe's PDFCreationSettings.pdf:
- Subsampling: A pixel in the center of the sample area replaces the entire area at the specified resolution. Subsampling is significantly faster than downsampling but results in images that are less smooth and continuous.
- Average downsampling: The pixels in a sample area are averaged, and the average pixel color replaces the entire area at the specified resolution.
- Bicubic downsampling: A weighted average is used to determine pixel color and usually yields better results than the simple averaging method of downsampling. This is the slowest but most precise method, resulting in the smoothest tonal gradations.