NEW Topaz upsampling product: A.I. Gigapixel

[plugsnpixels]

Thanks for your tests Victor. Surely we can expect Gigapixel to get better as time goes on; most software tends to.

[plugsnpixels]

I did some more tests today, this time with DSLR-sourced images and including ON1 Photo RAW in the mix.

[Bart_van_der_Wolf]

I did some more tests today, this time with DSLR-sourced images and including ON1 Photo RAW in the mix.

Hi,

Very interesting results, and assuming that the rightmost image is AIG, it demonstrates that the A.I. is adding 'plausible' detail where there is none to speak of. That's the main difference between traditional interpolation between existing pixels, and using a Neural network to replace pixels.

Cheers,
Bart

[plugsnpixels]

Thanks Bart, yes, Gigapixel is the right-most of each example.

Except for the cases where images are going to be presented for legal evidence and should be left "as-is" as much as possible, I can't see any reason not to be really excited about this new upsizing tech. I am really shocked at how much better the results seem to be when compared to traditional upsizing methods.

I would like to see some results from other LL members with better original photos and equipment than I have, and with other software I haven't tested yet.

[vjbelle]

Although looking at a file at 100% pixels tells something there is nothing like a print to really see how this software works.  I own both a GFX and a Phase 3100 and shot the same scene (front of my house that has tons of tiny plant detail) using same lens and adjusting distance to subject to account for focal length change.  The GFX file was upsampled to 40 inches on the wide side at 600dpi.  The 3100 file was also upsampled very slightly with preserve details 2.0 to 40 inches on the long side at 300dpi.  I printed to my ipf8400 using the print plugin at 600dpi.  That means that the GFX file never saw the Canon upsampling being a native 600dpi file.  The Phase 3100 was upsampled to 600dpi using the print plugin - which has been my print workflow. 

The final product is really kind of staggering.  I'm really fussy about print detail which is the reason I limit upsampling.  I would never upsample a GFX file beyond 40 inches unless the content could visually sustain the upsampling. Landscape detail already suffers at 40 inches - upon close inspection.  My benchmark is my Phase 3100 which can easily handle 40 inches.  The test crops were printed on Epson semi gloss photo paper which will easily show file anomalies.

To my surprise the GFX AI Giga upsampled file easily, visually, matched the detail of the Phase 3100 file.  This is with very close inspection when viewing at my EVS Executive Viewing Station.  Fine plant detail appears very natural and very sharp.  There is an overall smoothness to the print that is very pleasing.  You can't see any of this on a monitor - you have to print to see how this software works.  Again........ I'm stunned!  I could care less about the 20 minute wait as the results are worth it. 

To think that this technology is only going to get better is very exciting. 

Victor

[plugsnpixels]

Thanks Victor for your report! Can we come over and see your prints? ;-)

[vjbelle]

Any time you would like.....

Victor

[ErikKaffehr]

Hi,

No great surprise but still nice to hear about it. I got the impression that GFX has nice lenses. But, we need also discussing things like aperture, once we are past f/8, diffraction may be a limiting factor.

The GFX will need less stop down than larger formats like 54x41 mm, to keep same DoF.

Best regards
Erik

Although looking at a file at 100% pixels tells something there is nothing like a print to really see how this software works.  I own both a GFX and a Phase 3100 and shot the same scene (front of my house that has tons of tiny plant detail) using same lens and adjusting distance to subject to account for focal length change.  The GFX file was upsampled to 40 inches on the wide side at 600dpi.  The 3100 file was also upsampled very slightly with preserve details 2.0 to 40 inches on the long side at 300dpi.  I printed to my ipf8400 using the print plugin at 600dpi.  That means that the GFX file never saw the Canon upsampling being a native 600dpi file.  The Phase 3100 was upsampled to 600dpi using the print plugin - which has been my print workflow. 

The final product is really kind of staggering.  I'm really fussy about print detail which is the reason I limit upsampling.  I would never upsample a GFX file beyond 40 inches unless the content could visually sustain the upsampling. Landscape detail already suffers at 40 inches - upon close inspection.  My benchmark is my Phase 3100 which can easily handle 40 inches.  The test crops were printed on Epson semi gloss photo paper which will easily show file anomalies.

To my surprise the GFX AI Giga upsampled file easily visually matched the detail of the Phase 3100 file.  This is with very close inspection when viewing at my EVS Executive Viewing Station.  Fine plant detail appears very natural and very sharp.  There is an overall smoothness to the print that is very pleasing.  You can't see any of this on a monitor - you have to print to see how this software works.  Again........ I'm stunned!  I could care less about the 20 minute wait as the results are worth it. 

To think that this technology is only going to get better is very exciting. 

Victor

[plugsnpixels]

I received an email today asking why I hadn't also tested BenVista PhotoZoom Pro 7 against Topaz Gigapixel. Well, here you go ;-)

[vjbelle]

Hi,

No great surprise but still nice to hear about it. I got the impression that GFX has nice lenses. But, we need also discussing things like aperture, once we are past f/8, diffraction may be a limiting factor.

The GFX will need less stop down than larger formats like 54x41 mm, to keep same DoF.

Best regards
Erik

Hi Erik...... I think you are missing the point.  My test was shot with a 72mm Digitar - not a Fuji Lens.  This test was not about lenses but about upsampling software.  My Cambo is set up to accept the GFX and/or a DB.  That is how I am able to test both sensors with the same lens at the same aperture.  What is surprising to me is that, since I print large, I have been extremely concerned about upsampling and the results from the AI Giga have surpassed anything I have ever used.    The 3100 eliminated a lot of upsampling issues but even so I am so finicky that I don't even print 48 inches long without stitching two shifted files from my 3100.  To have a single 50mp file from the GFX equal a single 100mp file from my 3100 in print at 40 inches is no small feat.  I've used every upsampling software there is from ON1 perfect resize to Super Zoom to Adobe 'Preserve Details 2.0 and the print is the final judge and jury.  So far AI Giga is the real deal and the best there is..... only to get better.

Victor

[Paul2660]

Hi Victor,

How are you picking a dpi for output in Topaz?  I see the scale options, 200, 400 600 percent, but way to set a dpi?

thanks
Paul

[vjbelle]

Use inches not pixels for the width and other options will become available which includes 'dpi'.

Victor

[vjbelle]

I have run across one bug that I haven't been able to rectify.  If I use pixels as the width or height value then I cannot enter any more than 4 significant digest into either field.  The only way so far around this is to use inches in the 'fit to' value field.  I have sent Topaz a help request so will see if its universal or just me.  Has anyone else experienced this?

Victor

[Paul2660]

thanks Victor.

Just finished a test from a D850 image.  MacPro, 2 years old, Xeon 4 way, can't remember speed, 64GB of ram, and supposedly 2 radeon graphics cards.

Tried a 600%, and it took close to an hour to run, and when it finished the image was 4.9 GB, but CC could not "parse" the file and thus would not open it, never have seen such an error before in photoshop.  I wonder if it had to do with the fact the file was a tif and over 4GB? will Adobe not open a tif over 4GB, I know it won't save one that large.

Ran a 2nd one at 400%, this one came out to be a 73" x 110" image at 300 dpi, (where I tend to keep everything), I do print everything from Lightroom at either 240 or 360 dpi.    Did the same thing with Adobe CC preserve details 2. 

Overall just looking at 100% the Topaz file is better, Adobe shows a lot more strange noise and or mottling to the image, which is very clean at base output resolution of 300 in LR.  I don't understand why Topaz saves the file considerably smaller, 2.9GB and CC was over 4 GB and needed to be a PSB. 

I will try to post some samples a bit later, but it's always hard to get the point across IMO. 

Main issue for me now is time.  It still took the same MacPro about 45 minutes to run the uprez.  Not sure if some of this is due to the machine power settings, as it will always turn off the screen if left running for that length of time, and may be timing out also.  Need to check my power off/sleep settings.

Will work up a few more test shots in the next few days. Biggest issue for me is resources and processing time.  As Bart mentioned, Topaz is taking a huge amount of the system to run as even web browsing on that Macpro is much slower while the tool is running in the background.

Paul C

[plugsnpixels]

Thanks Paul, your tests remind me that a colleague at work has an iMac Pro which might be fun to test with Gigapixel. I also have access to 2013 Mac Pro's.

I did more onscreen tests today on the MBP, changing pixel mush into something usable.

600% upsampling of the lower left of this storefront image (shown here in reduced size for reference) as viewed in Photoshop at 100% (native pixels/Gigapixel).

[vjbelle]

Paul.... for some reason I get an error message from PS when opening the Giga file that says that the metadata is unreadable but the file still opens.  This is even after I tell Giga to include metadata.  I don't know if that is your issue. 

The reason I upsample to 600dpi is because I want to bypass any and all other driver/software upsampling.  I have always found this to yield the best results.  If I were printing to an Epson then I would upsample to 720dpi.

Today I took an old 1ds file taken back in 2003 and processed it in PS and upsampled in Giga to 30 inches at 600dpi.  I haven't printed it yet but the detail on screen at 33% pixels (which is a good screen visualization as to what the print will look like upon close inspection) is startling.  I would have never upsampled that file beyond 20 inches on the long side in the past.
 
I do hope that Topaz fixes the bug I mentioned.  I want to be able to enter pixel dimensions as I print to dimensions that are just a smidgeon more or less than standard (instead of 4X3 it may be 4.11X2.97, etc).  This is due to the fact that I museum wrap and stretcher bars are never perfectly sized.  I want the face of the frame to contain the image and the sides (depth of the stretcher bars) to be white.... its very difficult to accurately do and its rare that I get it perfect. I can get around all of this by using inches and printing slightly on the large side and then cropping but those are extra steps. 

Please post on your observations of an actual print. 

Victor

[Bart_van_der_Wolf]

thanks Victor.

Just finished a test from a D850 image.  MacPro, 2 years old, Xeon 4 way, can't remember speed, 64GB of ram, and supposedly 2 radeon graphics cards.

Tried a 600%, and it took close to an hour to run, and when it finished the image was 4.9 GB, but CC could not "parse" the file and thus would not open it, never have seen such an error before in photoshop.  I wonder if it had to do with the fact the file was a tif and over 4GB? will Adobe not open a tif over 4GB, I know it won't save one that large.

Hi Paul,

TIFF has a limit of 4GB, anything larger should be saved to either the BigTIFF or the PSB format.

I expect that AIG will get a future enhancement to save in such formats. I also expect that they may add (if users request it), a possibility to add noise or dithering at the final output dimensions. That would prevent some of the (too) smooth featureless regions (which risk posterization in 8-bit/channel output) at maximum scaling settings.

To reduce the risk of posterization as much as possible, one could tag the source image with the output profile before scaling which avoids an additional colorspace conversion with too few bits of precision, and that allows the image to be saved as an 8-bit/channel TIFF because it is the final data with no further editing expected.

Since upgrades are free for life, I expect it will only take some time before users are rewarded with these refinements.

Ran a 2nd one at 400%, this one came out to be a 73" x 110" image at 300 dpi, (where I tend to keep everything), I do print everything from Lightroom at either 240 or 360 dpi.    Did the same thing with Adobe CC preserve details 2.

I'd try 720 PPI (if the file size can be kept under 4GB), and once the preset is added (otherwise calculate the number of pixels required). This will produce even more accurate detail, and AIG will use every pixel it can to add detail.

Overall just looking at 100% the Topaz file is better, Adobe shows a lot more strange noise and or mottling to the image, which is very clean at base output resolution of 300 in LR.  I don't understand why Topaz saves the file considerably smaller, 2.9GB and CC was over 4 GB and needed to be a PSB.

Don't know why, but possibly the smoother image can be losslessly compressed further. 

Cheers,
Bart

[hubell]

Very interesting discussion. Thank you for the tests and the commentary. I have a few questions that seem relevant to the issue of file size and "best practices" with  respect to using AIG, given the issues about the speed of the program.
I print from an Epson 9900 printer and have upressed my files, which start at 360 ppi, to the desired size, leaving the file in 16 bit format and the "native" color space (Prophoto). I just feed the 16 bit file to the Photoshop print dialog box and pick the appropriate profile for the desired paper.
Should you upress to the desired print size with AIG at 720 ppi? Should you convert the 16 bit file to 8 bit before upressing? Should you convert the file to the desire paper profile before upressing?

[Paul2660]

Thanks Bart, I will try the 720 ppi but that well may again take the file past 4GB on a 400% uprez.

The single biggest issue for is time.  It took 1 hour to uprez a Fuji image (2 part pano) and about 45 to 40 minutes to uprez the D850 images. 

IMO a bit too long. 

Paul C

[Bart_van_der_Wolf]

Very interesting discussion. Thank you for the tests and the commentary. I have a few questions that seem relevant to the issue of file size and "best practices" with  respect to using AIG, given the issues about the speed of the program.
I print from an Epson 9900 printer and have upressed my files, which start at 360 ppi, to the desired size, leaving the file in 16 bit format and the "native" color space (Prophoto). I just feed the 16 bit file to the Photoshop print dialog box and pick the appropriate profile for the desired paper.
Should you upress to the desired print size with AIG at 720 ppi? Should you convert the 16 bit file to 8 bit before upressing? Should you convert the file to the desire paper profile before upressing?

Hi,

If the upsampled/scaled image has yet to undergo a color space conversion, e.g. from ProPhoto RGB to the output profile, I'd keep it in 16-bit/channel mode. After the conversion to the output profile, it can be made smaller by using 8-b/ch mode, even as a high-quality JPEG, because most printer driver pipelines are 8-b/ch.

If the image to be upsampled is already converted and tagged with the output profile/colorspace, then AIG will probably have no difficulty with scaling an 8-b/ch input file (JPG or TIFF) to an 8-b/ch output file.

If the final file size limit (4GB) is not exceeded, I'd try the highest native printer resolution, i.e. 600 or 720 PPI. Make sure to set the printer driver's options to accommodate the higher level of detail (i.e. "finest details" must be checked to on if an Epson is used, otherwise it will do a mediocre downsample to 360PPI).

How noticeable the difference will be, depends on the input image (subject & quality) and amount of upsampling. The 4x as many pixels will take longer to be calculated, so it's a trade-off which is only justified if the quality jump is visible enough.

Cheers,
Bart