Print Driver Upsampling or AI Gigapixel?

[Schewe]

From the literature I've consulted, the 300PP is stated as an estimate of the limit of human visual acuity, which of course must be a generalization that would Jeff has  suggested a range of resolutions one can send to the printer and still expect excellent outcomes; I don't think he had exotic resampling tools in mind when making those observations, but he should comment if he so wishes.

Read the article Andrew mentioned...that's still what I think. And to through a further wrench into the works, I'm pretty sure the Mac driver and Windows drivers behave differently on a range of issues except color management and profiles. So knowing EXACTLY what is going on in the print pipeline regarding resolution and interpolation is further problematic. In any event ANY interpolation is likely better in PS or LR or ACR or 3rd party apps than the friggin' print pipeline where at best it's a bilinear interpolation and at worst nearest neighbor (which really does suck).

And upsampling to get to 360/300 or 720/600 (Epson/Canon or HP) and THEN output sharpening is better than sending a non-standard resolution and hoping the print pipeline does ok.

[Doug Gray]

Read the article Andrew mentioned...that's still what I think. And to through a further wrench into the works, I'm pretty sure the Mac driver and Windows drivers behave differently on a range of issues except color management and profiles. So knowing EXACTLY what is going on in the print pipeline regarding resolution and interpolation is further problematic. In any event ANY interpolation is likely better in PS or LR or ACR or 3rd party apps than the friggin' print pipeline where at best it's a bilinear interpolation and at worst nearest neighbor (which really does suck).

And upsampling to get to 360/300 or 720/600 (Epson/Canon or HP) and THEN output sharpening is better than sending a non-standard resolution and hoping the print pipeline does ok.

Some time back I did a printer driver test with my 9800 and Win 10 x64. It's nearest neighbor. "At worst" indeed.

[JRSmit]

My personal experience is that Printing at 720 ppi andere fine details on given a visibly better result compared tot 360ppi.
Not only at closer viewing distance, also at longer distances, like onze would use to observe a print in a gallery.
In fact given the conversion to rather random placed dots of ink, i  do not see the pixels as such. Or the dots. What i see and experience is that Printing at 360ppi adds some softness, some vagueness , a sort of veil to the print. Also a feeling  that tonal changes are not so smooth. The perceived 3-dimensionality,if present on the image, is also impacted. For me it is about the experience when watching a print. And highest resolution does improve that. This is also the feedback i het from my customers.

[Bart_van_der_Wolf]

... In any event ANY interpolation is likely better in PS or LR or ACR or 3rd party apps than the friggin' print pipeline where at best it's a bilinear interpolation and at worst nearest neighbor (which really does suck).

And upsampling to get to 360/300 or 720/600 (Epson/Canon or HP) and THEN output sharpening is better than sending a non-standard resolution and hoping the print pipeline does ok.

Couldn't agree more.

Dedicated upsampling is a processing intensive operation (which e.g. Topaz A.I. Gigapixel takes to an extreme), and it's obvious that the printer driver pipeline does a worse job. And then we can add output sharpening after the resampling operation, something that the printer driver doesn't do by default either (and not sharpening Nearest Neighbor resampled data makes sense).

Cheers,
Bart

[Bart_van_der_Wolf]

My personal experience is that Printing at 720 ppi andere fine details on given a visibly better result compared tot 360ppi.
Not only at closer viewing distance, also at longer distances, like onze would use to observe a print in a gallery.

Hi Jan,

Although difficult to quantify (due to many variables), it is to be expected that boosting the resolution near the limiting resolution of the printer will also boost the MTF of the lower spatial frequencies. It's a bit like lifting a veil, the entire image acquires a bit more 'pop'. So having higher resolution and the ability to sharpen and/or boost local micro-contrast, has an impact on the overall image quality.

Topaz A.I. Gigapixel increases the real resolution beyond what is present in the source image, and thus lifts the veil by adding plausible high-resolution detail.

Cheers,
Bart

[JRSmit]

Hi Bart,

Thusfar i used for uprezzing the tool you and others crafted in 2014 here on lula. Also i would reduce the  sharpering to 0, and apply sharpering achter the uprez. Now gai looks like doing a better job since version 1.0.2 .
Not yet tested in prints. On my agenda for coming week.

[Mark D Segal]

............ In any event ANY interpolation is likely better in PS or LR or ACR or 3rd party apps than the friggin' print pipeline where at best it's a bilinear interpolation and at worst nearest neighbor (which really does suck).

And upsampling to get to 360/300 or 720/600 (Epson/Canon or HP) and THEN output sharpening is better than sending a non-standard resolution and hoping the print pipeline does ok.

Agreed - this is my standard working procedure as well.

Unlike for Mr. Smit, no advantage of 720 over 360 for photographs hits me in the face, and I like to believe I know what to look for in a photograph. Well, OK, things can look differently to different observers. What does your experience suggest?

[Schewe]

Agreed - this is my standard working procedure as well.
What does your experience suggest?

If the native resolution at print size is above 360/300 then it's better (and I've proved it for non-watercolor paper) to upsample to 720/600 and sharpen and print. If the native rez is below 360/300 then just stick to upsampling to 360/300 and don't bother with upsampling further.

All of this presupposes quality images properly adjusted with good lenses and no camera shake...YMMV if ya start with crap!

[Mark D Segal]

Jeff - puleese - I never start with "crap"; you should know that!?!?! (I hope)  :-)

Seriously - I agree - the GIGO principle assures that playing with these options won't do much for crappy input.

[vjbelle]

I'm curious who Topaz really targeted for the GIGA software.  Certainly not anyone in the high MP category - there aren't enough of us.  After using it, though, I am more than happy to receive the benefits.  I hope that they have become aware that even we/those users need this type of software development and advancement. 

Victor

[hubell]

If the native resolution at print size is above 360/300 then it's better (and I've proved it for non-watercolor paper) to upsample to 720/600 and sharpen and print. If the native rez is below 360/300 then just stick to upsampling to 360/300 and don't bother with upsampling further.

All of this presupposes quality images properly adjusted with good lenses and no camera shake...YMMV if ya start with crap!

If you start with a file that has a native resolution of less than 360 ppi at the desired print size,  you upress to 360 ppi using PS, LR or AIG, and then send the file to the Epson printer and check "Finest Detail," will the printer pipeline not upress from 360 to 720? If so, would you not be better off upressing in PS, LR or AIG to 720 dpi?
Thanks.

[Bart_van_der_Wolf]

If you start with a file that has a native resolution of less than 360 ppi at the desired print size,  you upress to 360 ppi using PS, LR or AIG, and then send the file to the Epson printer and check "Finest Detail," will the printer pipeline not upress from 360 to 720? If so, would you not be better off upressing in PS, LR or AIG to 720 dpi?

With AIG, yes, because AIG will attempt to add more resolution where there was none, to begin with. As long as you keep the resulting filesize below 4 GB (TIFF limitation).

Example, you have a narrow line, say a distant powerline or a twig or a crack in a surface, the other methods will make that line 6x as wide when scaling 600%. AIG will try to keep it narrower and sharp at just 4x times as wide, thus with higher resolution. This will not make a huge difference, but in some of the details it does show. And that is before any output sharpening is added. With more pixels, one can do a better job of sharpening for output.

I have an image example I'm willing to share.

• An original vertical panorama, that I downsampled to 30% of its original size:
  https://www.dropbox.com/s/kpps4axc14itsao/0920_Tr_30pct.jpg?dl=0
• A 300% AIG upscaled version (28.8 MB !) (that beats similar upsampling with Photozoom Pro and Perfect Resize):
  https://www.dropbox.com/s/jhr74lruhcss4cu/0920_Tr_30pct_Gpx110Mod300pct.tiff?dl=0 (NOTE: Updated to AIG Version 1.1.0, with moderate noise and blur reduction)
• A 600% AIG upscaled version (93.5MB !):
  https://www.dropbox.com/s/wp6ji82ab376wtl/0920_Tr_30pct_Gpx110Mod600pct.tiff?dl=0 (NOTE: AIG Version 1.1.0, with moderate noise and blur reduction)

Now, before you start inspecting the level of detail, make sure to zoom out to the printed dimensions, so use:
100 * Display_PPI / Printer_PPI
as the display zoom percentage to watch it at the printed output size (@600 PPI or 720 PPI).

There will hardly be a visual difference, because the display pixels are the same size. But now, realize that the 600/720PPI print will have almost twice the resolution of what is needed to satisfy our average visual acuity. It will be sharper.

An excellent print can be made (almost as good as when starting with the full-size original, not this downsampled version of it). The downsampling did lose some detail (because it became smaller than 1 pixel), but AIG did a fine job of restoring the impression of detail.

Cheers,
Bart

[Malcolm Payne]

I had an email from Ddisoftware yesterday, the authors of Qimage, with a link to a video directly comparing the output of Qimage and AI Gigapixel, which raises some interesting points: https://youtu.be/LGWEyG4DUOM

He actually recommends AIG in some instances, though he highlights some artefacts in its processing that might be counter-indicative. I haven't yet found comparable artefacts in my own tests of AIG, though on different subject matter - it would be interesting to know if anyone else has found anything similar.

Malcolm

[Bart_van_der_Wolf]

I had an email from Ddisoftware yesterday, the authors of Qimage, with a link to a video directly comparing the output of Qimage and AI Gigapixel, which raises some interesting points: https://youtu.be/LGWEyG4DUOM

He actually recommends AIG in some instances, though he highlights some artefacts in its processing that might be counter-indicative. I haven't yet found comparable artefacts in my own tests of AIG, though on different subject matter - it would be interesting to know if anyone else has found anything similar.

Hi Malcolm,

Thanks for that link. Mike's comments in that video are correct, and he explains the differences/trade-offs quite well.

Unfortunately for him, good for us, Topaz Labs have just issued an update (to probably a newer version than what Mike used). The users now have a choice of how strong the noise and Blur reduction must be (None, Moderate, or Strong).

I just ran the newer version on the image of the ship I posted before, and even in the Moderate setting, the results were visibly crisper with more subtle low contrast detail than in the Version 1.0.2. So Mike's results might also have been a bit different, had he used this later version of the software.

I have yet to compare it with the noise and blur reduction set to None, but Moderate is already an improvement for low noise image input compared to the previous version. It's good to have options.

Cheers,
Bart