NEW Topaz upsampling product: A.I. Gigapixel

[Bart_van_der_Wolf]

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.

Hi Paul,

While the folks at Topaz Labs will try to improve the required processing time, it will remain a very processor intensive operation (CPU+FPU+GPU). Even on my not supported video card, the GPU processing will frequently hit a 100% load. So it will never be a fast procedure for large files on current hardware. Check out the processor load to see which component is the bottleneck. Only with faster hardware can we expect to shave off waiting time.

That's why AIG does batch processing and runs as a separate application, so as to not block other processes (although they will slow down). If the computer is set to not go to sleep, it can run overnight.

Cheers,
Bart

[hubell]

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

Thanks, Bart.
Does AIG do any sharpening of the file as part of the upressing process? Hopefully not.

[vjbelle]

I always thought they did a small amount of sharpening but can't confirm.  As has been covered by Bart I always have in the past upsampled the image to be printed in my favorite program to desired DPI to avoid any other upsampling drivers/software on the printer side.  This was before I started to use the print plugin for my Canon printer which upscales to 600DPI which I started to use exclusively at that point.  In other words I would upsample only to 300DPI.  Now with Giga I just let the program upsample to my desired print size (40 inches minimum) at 600DPI to avoid even the Canon software.  The printed results have surpased any other workflow I have used.

This software will more than likely become a standard part of my workflow as any file that I want to print has to upsampled.  That includes anything (even stiched) that comes out of my 3100.  Phase doesn't have enough pixels now (minimum 24000 on the long side) nor will they anytime soon. 

Victor

[Bart_van_der_Wolf]

Thanks, Bart.
Does AIG do any sharpening of the file as part of the upressing process? Hopefully not.

Hi,

I don't think we can call it sharpening in the normal sense, and I do not see any signs of sharpening halo or ringing either. AIG is actually trying to restore the original sharpness that an image of that output size would have, had it been shot at that size.

That's how Neural Networks function, after being trained on a small version of the scene and on the full-size version of the same scene. It then learns how to replace interpolated pixels with the actual pixels that should be there. Given the possible variety of image content that it may be confronted with, it has to be trained on many different types of images.

This also means that it sometimes guesses wrong when dealing with other images than those it was trained on. This is what the folks at Topaz Labs are working on to improve, but I've read that it can take a week of additional training before they can see if a tweak gives better results or not.

Cheers,
Bart

[Paul2660]

Well, I tried a few on this Mac Pro at 720 dpi and the tool aborts about 3/4 of the way through.  Not sure why, no error.

But still making progress.

What I have realized is the "enhanced" setting pulls does add a bit of smoothness, and pulls out all the noise in areas like sky.  Holds up good edges and still has fine details. 

Will go ahead and add this to my tools, and do hope they get a better handle on the speed.

Paul C

[ErikKaffehr]

Hi Bart,

From what I have seen is that the algorithm segments detail and than renders those segments artificially.

Best regards
Erik

Hi,

I don't think we can call it sharpening in the normal sense, and I do not see any signs of sharpening halo or ringing either. AIG is actually trying to restore the original sharpness that an image of that output size would have, had it been shot at that size.

That's how Neural Networks function, after being trained on a small version of the scene and on the full-size version of the same scene. It then learns how to replace interpolated pixels with the actual pixels that should be there. Given the possible variety of image content that it may be confronted with, it has to be trained on many different types of images.

This also means that it sometimes guesses wrong when dealing with other images than those it was trained on. This is what the folks at Topaz Labs are working on to improve, but I've read that it can take a week of additional training before they can see if a tweak gives better results or not.

Cheers,
Bart

[walter.sk]

Tried AI Gigapixel on my laptop (Win10 64bit, 16GB RAM Geforce GTX 960M 4G) and got very slow but good results with AI G, upsampling 5184x3188 pixel files to 3X or 4X. I'm impressed enough with the quality, and when I get home I'd like to use AI G on my desktop  but I am limited by budget issues.  Here's the story:

I have an ancient Alienware Area 51 desktop running Win10 64B.  It has an i-7 4 core CPU, is maxed out at 12GB DDR-3 RAM, and I'm using 2 Geforce GTX 460 DDR-5 video cards, 1G each.  I am a heavy  user of LR Classic CC and PSCC, and have been happy with the old computer:  I use 6 internal hard drives (4 for images) as well as a bunch of external drives for backup.  The computer also has a 1500W power supply and liquid cooling, and it would cost much more than I can afford to replace it.  It runs just fine for my purposes.

The question:  The video cards are only 1G, are not used in parallel, and are on the AI G Not Supported list. The video cards are 256-bit, PCI-E 2.0 x16, width:dual slot. 

1) Am I right in thinking that I cannot use the newer PCI-E 3.0 cards (the computer was made in 2009)?
2) If I can find video cards that will work with the PCI-E2 specs, should I get one card with as much VRAM as possible?
3) Or is the PCI-E spec dependent on the video card rather than the computer slot, in which case I could get a PCI-E 3 card?

[Bart_van_der_Wolf]

Tried AI Gigapixel on my laptop (Win10 64bit, 16GB RAM Geforce GTX 960M 4G) and got very slow but good results with AI G, upsampling 5184x3188 pixel files to 3X or 4X. I'm impressed enough with the quality, and when I get home I'd like to use AI G on my desktop  but I am limited by budget issues.  Here's the story:

I have an ancient Alienware Area 51 desktop running Win10 64B.  It has an i-7 4 core CPU, is maxed out at 12GB DDR-3 RAM, and I'm using 2 Geforce GTX 460 DDR-5 video cards, 1G each.  I am a heavy  user of LR Classic CC and PSCC, and have been happy with the old computer:  I use 6 internal hard drives (4 for images) as well as a bunch of external drives for backup.  The computer also has a 1500W power supply and liquid cooling, and it would cost much more than I can afford to replace it.  It runs just fine for my purposes.

The question:  The video cards are only 1G, are not used in parallel, and are on the AI G Not Supported list. The video cards are 256-bit, PCI-E 2.0 x16, width:dual slot. 

1) Am I right in thinking that I cannot use the newer PCI-E 3.0 cards (the computer was made in 2009)?
2) If I can find video cards that will work with the PCI-E2 specs, should I get one card with as much VRAM as possible?
3) Or is the PCI-E spec dependent on the video card rather than the computer slot, in which case I could get a PCI-E 3 card?

Hi,

Most motherboards and computers manufactured in 2013 or later probably support PCI Express v3.0. So your 2009 desktop's motherboard probably doesn't support PCI-E. Generally, PCI Express refers to the actual expansion slots on the motherboard that accept PCIe-based expansion cards and to the types of expansion cards themselves.

PCI-Express 3.0 has already been superseded by 4.0, and 5.0 is expected to be released in 2019.

So the cheapest short-term solution woulld be to find a compatible video card for your motherboard's version of PCI. However, glancing at the Topaz Labs requirements page, it looks like it that the compatible desktop graphics cards must be at least PCI-E 3.0 versions.

So it looks like your laptop is going to get busy.

Cheers,
Bart

[walter.sk]

PCI-Express 3.0 has already been superseded by 4.0, and 5.0 is expected to be released in 2019.

So the cheapest short-term solution woulld be to find a compatible video card for your motherboard's version of PCI. However, glancing at the Topaz Labs requirements page, it looks like it that the compatible desktop graphics cards must be at least PCI-E 3.0 versions.

So it looks like your laptop is going to get busy.

So I feared.  Thanks for the reply.  At 77 years old, I'm getting used to this pattern.  Gotta win the lottery, I guess.   

[BrianWJH]

So I feared.  Thanks for the reply.  At 77 years old, I'm getting used to this pattern.  Gotta win the lottery, I guess.   

Should be compatible, PCIe Gen3 is supported on 3rd Generation of Intel® Core™ i5 and Core™ i7 CPUs

Brian.

[Malcolm Payne]

@walter.sk:

PCI-e 3.0 cards are backwards compatible, so if your existing cards are PCI-e 2.0 you should be OK with any recent graphics card on Topaz Labs compatibility list. I'm running a Quadro P4000 (PCI-e 3.0) on a vintage 2010 PC with a 1st gen i7-980x CPU on an Asus P6X58D-E m/b (PCI-e 2.0) with 12G RAM, and it runs AI Gigapixel just fine, with fairly decent processing speeds. If you don't need the workstation and 10-bit features of the Quadro, a recent gaming card will probably be faster still, and certainly cheaper.

No crashes or hangs so far even on large input files and enlargements, just one glitch where Photoshop complained about an incorrectly terminated file error but then opened it without any further problems.

FWIW, I've tested Gigapixel on a number of images (only on-screen so far), including a photo for reproduction of an oil painting with heavy and highly-detailed impasto, and have been seriously impressed. The latter showed no visible loss or degradation of detail at 200% increase; at 400% some of the fine texture appeared a little 'plastic' but comparison with a magnified view of the original showed the structure of the detail to be similar, and I suspect it will probably print better than it looks.

On landscapes with distant foliage, grass or similar fine detail there was some tendency for this to appear a little 'wormy' at 400%+, but again close examination showed this to be generally comparable to the magnified original, so the artefact is probably mostly a function of the raw conversion of fine detail at the resolution limit of the camera and lens rather than anything introduced by Gigapixel.

My previous 'goto' for difficult enlargements has been Photozoom Pro 6.0, but Gigapixel appears considerably better so far.

Malcolm

[walter.sk]

@walter.sk:
PCI-e 3.0 cards are backwards compatible, so if your existing cards are PCI-e 2.0 you should be OK with any recent graphics card on Topaz Labs compatibility list. I'm running a Quadro P4000 (PCI-e 3.0) on a vintage 2010 PC with a 1st gen i7-980x CPU on an Asus P6X58D-E m/b (PCI-e 2.0) with 12G RAM, and it runs AI Gigapixel just fine, with fairly decent processing speeds. If you don't need the workstation and 10-bit features of the Quadro, a recent gaming card will probably be faster still, and certainly cheaper.
Malcolm

That sounds more hopeful.  I think my CPU is a gen-1 i7-980 as well, but I can't check it for another few days.  From what I've observed, AI G is really impressive.

[Bart_van_der_Wolf]

@walter.sk:

PCI-e 3.0 cards are backwards compatible, so if your existing cards are PCI-e 2.0 you should be OK with any recent graphics card on Topaz Labs compatibility list. I'm running a Quadro P4000 (PCI-e 3.0) on a vintage 2010 PC with a 1st gen i7-980x CPU on an Asus P6X58D-E m/b (PCI-e 2.0) with 12G RAM, and it runs AI Gigapixel just fine, with fairly decent processing speeds. If you don't need the workstation and 10-bit features of the Quadro, a recent gaming card will probably be faster still, and certainly cheaper.

Hi Malcolm,

That's good to know (I didn't, so thanks for adding that info). I did read that a conventional PCI-card doesn't fit in a PCI Express-slot, so I assumed it to be the case the other way around as well. All I know is that the card's bus and the motherboard's slots need to match, and that PCI-E comes in different lengths, 1x - 16x.

FWIW, I've tested Gigapixel on a number of images (only on-screen so far), including a photo for reproduction of an oil painting with heavy and highly-detailed impasto, and have been seriously impressed. The latter showed no visible loss or degradation of detail at 200% increase; at 400% some of the fine texture appeared a little 'plastic' but comparison with a magnified view of the original showed the structure of the detail to be similar, and I suspect it will probably print better than it looks.

On landscapes with distant foliage, grass or similar fine detail there was some tendency for this to appear a little 'wormy' at 400%+, but again close examination showed this to be generally comparable to the magnified original, so the artefact is probably mostly a function of the raw conversion of fine detail at the resolution limit of the camera and lens rather than anything introduced by Gigapixel.

My previous 'goto' for difficult enlargements has been Photozoom Pro 6.0, but Gigapixel appears considerably better so far.

My impressions are the same, although I'm somewhat handicapped with non-compatible hardware. Photozoom Pro 6 did better than most others, until Gigapixel was released. It looks like we have a new King. The King is dead, long live the King.

Close inspection of the file on display will not accurately tell how it looks in print. Print usually looks less artificial.
 
Cheers,
Bart

[Malcolm Payne]

Hi Bart,

PCI and PCI Express are not compatible either physically or electronically. The PCI Express slots on my Asus m/b are PCI Express 2.0 x 16, while the Quadro P4000 is PCI Express 3.0 x 16; Walter's GTX 460 cards are PCI Express 2.0 x 16, so he should be OK with any PCI-e 3.0 (or higher?) replacement that satisfies the A.I. G minimum specs or better.

I'll try a test print or two from A.I. G when I get a moment and report back. I can also run a few timings if anyone is interested in the performance on my gear; despite the older CPU and relatively limited RAM it seems to hold up pretty well compared to other reported results, with processing times generally ITRO 10-20 min or so for enlargements from D850 files. No doubt the Quadro P4000 compensates for the other deficiencies in the system; it wasn't cheap but has usefully extended the lifetime of the build for several years more at least (my other primary uses for the Quadro are Capture One processing and 10-bit Photoshop display for art repro).

I'm wary of potentially using A.I. G for repro work at this stage apart from special use cases, simply because it is said to generate 'plausible' rather than actual detail in enlargements, but I've seen nothing yet far that actively disqualifies it from consideration for that purpose. For anything else that is not so reliant on the absolute integrity of the result, it looks to have a lot going for it.

Cheers,

Malcolm

[Bart_van_der_Wolf]

Hi Bart,

PCI and PCI Express are not compatible either physically or electronically. The PCI Express slots on my Asus m/b are PCI Express 2.0 x 16, while the Quadro P4000 is PCI Express 3.0 x 16; Walter's GTX 460 cards are PCI Express 2.0 x 16, so he should be OK with any PCI-e 3.0 (or higher?) replacement that satisfies the A.I. G minimum specs or better.

Ah, now I get it. So his older motherboard apparently was already a PCI Express (2.0). Then things look more promising.

I'll try a test print or two from A.I. G when I get a moment and report back. I can also run a few timings if anyone is interested in the performance on my gear; despite the older CPU and relatively limited RAM it seems to hold up pretty well compared to other reported results, with processing times generally ITRO 10-20 min or so for enlargements from D850 files. No doubt the Quadro P4000 compensates for the other deficiencies in the system; it wasn't cheap but has usefully extended the lifetime of the build for several years more at least (my other primary uses for the Quadro are Capture One processing and 10-bit Photoshop display for art repro).

Would be interesting, because many people report slow speeds, but they often don't mention with which specific hardware. Also, it might be useful to note the maximum % of the "Highest utilization across all GPU engines" that's reported in the Task Manager. Mine frequently hits the 100% ceiling, suggesting that with faster GPUs one would reduce the total time required. And maybe the Quadro architecture has also other benefits for this specific use case.

I'm wary of potentially using A.I. G for repro work at this stage apart from special use cases, simply because it is said to generate 'plausible' rather than actual detail in enlargements, but I've seen nothing yet far that actively disqualifies it from consideration for that purpose. For anything else that is not so reliant on the absolute integrity of the result, it looks to have a lot going for it.

Yes, although 'regular' interpolation also produces fake/plausible information. My expectation is that AIG on average gets closer to the 'ground truth' than the more blurry methods, especially on the type of detail that human vision relies on, i.e. edges and lines. The only way to be absolutely sure of correct detail is by shooting at a larger magnification (closer up, or with longer focal length), which may then require stitching for a wide enough Field of View.

Cheers,
Bart

[Malcolm Payne]

Hi Bart,

I'm inherently a purist for repro work and I do regularly stitch for larger artworks; routinely a 3-way flat stitch with a PC-E Micro Nikkor 85mm shift lens and sometimes a multi-column version of same for the largest. PTGui Pro (now v 11) does a great job on the stitching. That and a good RIP takes care of most things.

But there are times when something like A.I. G could potentially be a very useful addition to the toolbox, especially for extreme enlargements or when clients need a print from their own historic files or transparencies where the original work can't be re-photographed to the necessary standards. In such cases any potential minor inaccuracies in reproduction, while not desirable, are usually an acceptable trade-off for getting a print at the required size and of a satisfactory quality.

Parenthetically, and not wanting to derail this thread, I've experimented with regular panoramic stitching for even higher resolution but haven't found the results satisfactory for artwork. Possibly it's a function of the lens used or I haven't nailed the right technique yet, as I know it has been used for large digitisation projects.

Cheers,

Malcolm

[vjbelle]

Would be interesting, because many people report slow speeds, but they often don't mention with which specific hardware. Also, it might be useful to note the maximum % of the "Highest utilization across all GPU engines" that's reported in the Task Manager. Mine frequently hits the 100% ceiling, suggesting that with faster GPUs one would reduce the total time required. And maybe the Quadro architecture has also other benefits for this specific use case.

Cheers,
Bart

Hi Bart..... I have never seen 100% GPU usage in Task Manager but do hit 75% but only for milliseconds.  This is with a 1070 series card.  My times are routinely 20 minutes for upsizing a 50mp image to 40 inches on the long side at 600dpi which equals 24000 pixels.  The size of the original file has a lot of influence on times. I've toyed with upgrading the GPU to the new 2080/70 series but really don't know if it would be worth it. 

Victor

[Bart_van_der_Wolf]

Hi Bart..... I have never seen 100% GPU usage in Task Manager but do hit 75% but only for milliseconds.  This is with a 1070 series card.  My times are routinely 20 minutes for upsizing a 50mp image to 40 inches on the long side at 600dpi which equals 24000 pixels.  The size of the original file has a lot of influence on times. I've toyed with upgrading the GPU to the new 2080/70 series but really don't know if it would be worth it.

Hi Victor,

If the GPU utilization doesn't hit the 100% ceiling, then there is some reserve left and a different GPU would probably only make a significant change if it uses much higher clock frequencies, or faster memory and bus speeds. Of course, there may also be other architecture improvements possible, but at least the GPU utilization is not determining the limit in your case.

Cheers,
Bart

[Paul Roark]

This does appear to be interesting software.  I just posted a comparison/example at http://paulroark.com/AI-Gigapixel-sample.JPG .   For large files (more than a single Sony a7rii frame) I ran into insufficient RAM and other issues.  (I'm on a Microsoft Surface Studio.)  Also, I noted on an old photo I'd copied for someone, there was no improvement in quality -- garbage in, garbage out.

I think some of the smoothed areas in the sample I worked on were too smooth.  I don't know if there is any adjustment to stop this over-smoothing, which, in effect, eliminated some image content and made the output look a little too much like plastic.  I think I'll try using standard bicubic on the original, put is as a lower layer, and then seeing if revealing a little of the original helps the "over-smoothed" areas.

(Note that the Sony I used for the posted sample has the KolariVision Ultra Thin cover glass.  Leica M wide angles do not work well on Sony cameras with the OEM thick cover glass.)

Paul
www.PaulRoark.com

[Bart_van_der_Wolf]

This does appear to be interesting software.  I just posted a comparison/example at http://paulroark.com/AI-Gigapixel-sample.JPG .   For large files (more than a single Sony a7rii frame) I ran into insufficient RAM and other issues.  (I'm on a Microsoft Surface Studio.)  Also, I noted on an old photo I'd copied for someone, there was no improvement in quality -- garbage in, garbage out.

Hi Paul,

Absolutely true, garbage only becomes bigger. But on the other hand, quality also increases. So it pays to use good quality input, properly capture sharpened (without halos and other artifacts), and Clarity applied for effective local contrast enhancements. Noise reduction, I'd be careful about that, because there is information mixed in with the noise. And A.I. Gigapixel effectively suppresses meaningless noise in favor of meaningful detail.

I think some of the smoothed areas in the sample I worked on were too smooth.

Judging from the linked example you posted, I'd say that actual detail with noise was pretty well preserved as only detail, and noise without detail was traded for lack of detail.

I don't know if there is any adjustment to stop this over-smoothing, which, in effect, eliminated some image content and made the output look a little too much like plastic.  I think I'll try using standard bicubic on the original, put is as a lower layer, and then seeing if revealing a little of the original helps the "over-smoothed" areas.

Yes, one could consider a low opacity blend of some of the 'detail' as rendered by e.g. Photoshop or similar, or adding a little amount of fine grain that can fool our eyes into seeing non-existent detail. That is currently not a built-in feature that AIG has, but who knows, they might add it in the future.

Cheers,
Bart