Luminous Landscape Forum
Raw & Post Processing, Printing => Digital Image Processing => Topic started by: plugsnpixels on September 27, 2018, 12:59:05 pm
-
As expected, Topaz is continuing to improve A.I. Gigapixel with the new version (2.0.0) making a huge breakthrough in performance. See my tests here (http://"http://plugsandpixels.com/blog/released-today-topaz-a-i-gigapixel-2-0-5x-faster/").
Topaz says, "Using GPU mode is now 5x faster! With some optimizations to the GPU processing in A.I. Gigapixel, we've improved the speed of processing by around 5x on full-size images (it scales based on the size of your image, so a small image might be 3x) with the same 1-2 million calculations per pixel and the same amazing results."
People who have tried Gigapixel have mentioned the unsurpassed quality of the results, but also the loooooong time it took to get them. So this new speed boost is welcome news.
New Features:
•Improved processing speed increased up to 5X
Bug Fixes:
•Made Cancel terminate processing immediate and clears the image list
•Note: Topaz is not promoting a coupon code or discount for A.I. Gigapixel but see my blog post (http://"http://plugsandpixels.com/blog/released-today-topaz-a-i-gigapixel-2-0-5x-faster/") to save 15% off the retail price (this discount also applies to ALL Topaz products). The latest update is FREE for existing users, as per Topaz's longstanding policy for all of its software.
-
My older version upgraded automatically last night when I used it. The upgrade speed is very apparent and welcomed.
While we're talking adding features, any chance of removal of Jpeg artifacts from the original file?
Paul
www.PaulRoark.com
-
While we're talking adding features, any chance of removal of Jpeg artifacts from the original file?
Does the software work from the RAW file? Can you choose the output format, e.g. 16-bit TIFF?
The way to avoid jpg artifacts is not to shoot jpgs, or if you must, choose the uncompressed kind. It seems silly to use this kinds of software if you are just trying to restore data you have thrown away by using compressed jpgs to begin with.
-
Does the software work from the RAW file? Can you choose the output format, e.g. 16-bit TIFF?
Hi,
Yes, it does convert Raws, but by doing that one misses the opportunity to create a better source image (e.g. Chromatic Aberrations) before scaling.
The way to avoid jpg artifacts is not to shoot jpgs, or if you must, choose the uncompressed kind. It seems silly to use this kinds of software if you are just trying to restore data you have thrown away by using compressed jpgs to begin with.
Yes, but sometimes a JPEG is all one has/gets (e.g. most smartphone images, or from a Social Network).
Cheers,
Bart
-
My older version upgraded automatically last night when I used it. The upgrade speed is very apparent and welcomed.
Hi Paul,
Good to know. My graphics card is not supported, so my scaling is mostly done with the CPU, and that speed is not all that much different.
While we're talking adding features, any chance of removal of Jpeg artifacts from the original file?
The "reduce Noise and Blur" selection also changes the way that JPEG artifacts are handled.
One also has the option to use multiple rescalings, e.g. one on "Strong" (to suppress the gravest of artifacts) and layer that in Photoshop with a version that was scaled on "None", and then blend them with a mask that transitions between the 8x8 pixel block edges.
Or, maybe better, is to preprocess the JPEG e.g. with Topaz DeJPEG (https://topazlabs.com/dejpeg/).
Cheers,
Bart
-
I was notified that it was updated, now to v2.0. Unfortunately, my trial period from the original download of v1.0 has expired so I cannot test it. I'm not sure I understand why I cannot - it is a completely new version, so I would expect that I could get another 30 days out of a trial. Oh well.
kirk
-
I was notified that it was updated, now to v2.0. Unfortunately, my trial period from the original download of v1.0 has expired so I cannot test it. I'm not sure I understand why I cannot - it is a completely new version, so I would expect that I could get another 30 days out of a trial. Oh well.
Hi Kirk,
I do not know the exact reason why they changed the major version number within some 2 months, but it has 'only' changed processing speed (which presumably required a significant enough architecture change to justify the new version numbering). But then for Topaz products, there are no upgrade fees for life, for previous buyers, so maybe they thought it was too soon for another 30 days trial. The purchasing risk is low, and the potential benefits (as promised) are already materializing. That's why I decided to take the plunge, even with a non-supported graphics chart.
The basic functionality has not changed, so maybe they considered that that was not enough to warrant a new 30 day (re)evaluation term. I do understand that a few users want to reassess the processing speed to see if the processing time fits in their specific workflow restrictions.
All I can say is that the output quality surpasses anything that I used/tried before (and that were many solutions), so I have no regrets. Others may have other considerations.
Cheers,
Bart
-
As expected, Topaz is continuing to improve A.I. Gigapixel with the new version (2.0.0) making a huge breakthrough in performance. See my tests here (http://"http://plugsandpixels.com/blog/released-today-topaz-a-i-gigapixel-2-0-5x-faster/").
Topaz says, "Using GPU mode is now 5x faster! With some optimizations to the GPU processing in A.I. Gigapixel, we've improved the speed of processing by around 5x on full-size images (it scales based on the size of your image, so a small image might be 3x) with the same 1-2 million calculations per pixel and the same amazing results."
People who have tried Gigapixel have mentioned the unsurpassed quality of the results, but also the loooooong time it took to get them. So this new speed boost is welcome news.
I'm impressed with the AI results in extrapolating based on deep knowledge of real images. Quite impressive.
-
Thanks Bart for your comments. I too was going to mention Topaz DeJPEG (http://plugsandpixels.com/dejpeg.html) (one of their classic products), which would help prep the image going into AI Gigapixel. I also did some experiments using A.I. Clear (http://plugsandpixels.com/blog/topaz-a-i-clear-a-i-gigapixel-vs-film-grain-they-win/) to prep noisy film-based images, but this might apply to bad JPEGs as well.
kirkt, I wrote Topaz about the trial reset situation and they said: "We can't actually reset [Gigapixel] trials anymore as they're all the same product. However, if they purchase and are unsatisfied they can refund their purchase within 15 days, no questions asked."
Hope that helps.
-
However, if they purchase and are unsatisfied they can refund their purchase within 15 days, no questions asked."
Good point. I was offered a refund (for the unsupported Graphics Card, although I could have known before purchasing) as well, but since the CPU/GPU selection was added I had no reason to pursue that option.
Cheers,
Bart
-
Thank Bart, that CPU option was a nice surprise.
What hardware are you using again? So far I haven't used Gigapixel on anything newer than 2013 Macs and it's been fine with GPU.
-
Thanks for the explanation and advice - sounds like I will just bite the bullet and purchase it and see how it evolves.
By the way, as more and more products in the image processing arena are released with these machine learning approaches, consider watching a couple of incredibly well made video series regarding:
Machine learning: "Learning to see" - https://www.youtube.com/watch?v=i8D90DkCLhI&list=PLiaHhY2iBX9ihLasvE8BKnS2Xg8AhY6iV
Neural Networks: "neural Networks Demystified" - https://www.youtube.com/watch?v=bxe2T-V8XRs&list=PLiaHhY2iBX9hdHaRr6b7XevZtgZRa1PoU
These are some of the most well-made and communicative videos you will find on a topic. The combination of the clear and well articulated narrative, along with the well-produced and succinct graphical presentation of the information is top notch.
Kirk
-
Great videos explaining a difficult subject!
-
The newest version still takes some time while Iridient takes about 10 seconds to do the same thing. Gigapixel has slightly better detail but a slight sharpening
on the Iridient version would make it just about equal. Here's a 50% view of the 2 processed at 250% upscale from a file from the Sony a7r3.
-
Thanks Wolfman. Is this with Iridient Transformer? Doesn't seem like it, though Developer doesn't seem to have upsampling ability (??).
-
It is Iridient Developer. It has many scaling methods. I have not tested all of them yet. Many to choose from.
-
Here's what I see (first time checking out the product). Version 3.2.3 demo.
-
On the upper right it says out...you click on that and you get these windows.
-
Oh I see, the settings are on the way out the door! "Normally" that gets set up first, but OK.
-
It is Iridient Developer. It has many scaling methods. I have not tested all of them yet. Many to choose from.
Hi,
Yes, these are common interpolation methods. The Lanczos methods are popular, but they produce ringing artifacts that need to be suppressed. Topaz Gigapixel doesn't produce ringing artifacts or halos. AIG also adds detail that's smaller than the original detail times the scaling factor.
Your example was at 50% zoom scale which halves the resolution and somewhat hides the traditional artifacts and noise magnification.
Cheers,
Bart
-
I looked at it at 50% because that is the % that I look at a file for printing and checking sharpening.
-
100% view.
-
Bart,
I was told by a highly respected landscape photographer and processing guru who works closely with Art Printers Of Nevada that the owner of Nevada Printers is always researching & testing the latest cutting edge upscaling techniques for large print making and he says that the Lanczos algorithm in Raw Therapee is the best upscaling method at this time. Granted I used it in Iridient & not Raw Therapee.
I don't know if there is a difference between those 2.
-
Does anyone here use RawTherapee? I've heard of it and keep downloading it and tucking it away but never used it, until right now when I decided to see what Wolfman was talking about. And I can't figure out how it works (short of having to go read something...).
I finally found the Transform>Resize area but was unable to input my desired Resize parameters (is the pre-populated field pulling from the existing image size? Is it in pixels (probably) or cm or inches? The input fields maxed out at 3564 whatevers and I wanted something higher.
I went with the numbers it allowed. Now where do you go to invoke the process and output the results? I see there's a queue. I must have done something because now I've got .pp3 RawTherapee Profile Data files in my image folder for whatever reason but no enlarged images (unless the same pixel dimensions at twice the file size means anything).
Plus the interface is slow. Maybe it's still processing my request? There's no progress bar active so who knows.
Anyway, some software is just not fun to use... ;-)
-
Bart,
I was told by a highly respected landscape photographer and processing guru who works closely with Art Printers Of Nevada that the owner of Nevada Printers is always researching & testing the latest cutting edge upscaling techniques for large print making and he says that the Lanczos algorithm in Raw Therapee is the best upscaling method at this time. Granted I used it in Iridient & not Raw Therapee.
I don't know if there is a difference between those 2.
Hi,
Yes, in landscape resampling with a Lanczos windowing filter is quite useful, because the ringing artifacts are harder to spot than in e.g. architecture or product shots.
I've intensively studied the various interpolation algorithms and methods that are around, and I contributed an open source resampler based on ImageMagick here on LuLa (https://forum.luminous-landscape.com/index.php?topic=91754.msg746273#msg746273). It uses a mix of upsampling settings at various gamma settings which are then blended together to avoid those resampling artifacts as best possible.
Presumably, some form of this technique found its way into the Capture One Raw converter, a few versions ago. That also benefitted the quality of lens distortion corrections, rotations, and keystone corrections.
But what all these traditional interpolation methods have in common, is that they mathematically create smooth transitions between existing pixels. The Lanczos family of filters performs very good, but with caveats that need to be addressed/corrected (which not all software does because it slows down the procedure).
Topaz A.I. Gigapixel works in a different way, in that it has learned from examples what that detail between existing pixels probably actually looked like at high resolution and uses that to fill in the blanks instead of using a weighted average. This allows it to sometimes create actual detail with higher resolution than the original capture.
Cheers,
Bart
-
I'm impressed with the AI results in extrapolating based on deep knowledge of real images. Quite impressive.
Your Link is not working.
-
Is your computer too old to run Gigapixel?
I just tested on a 2008 MacBook Pro running Mojave (yes!), 4 gigs RAM, SSD. It works!
-
I just tried version 2.0 for the first time, and the speed increase is impressive. An image that took 19-20 minutes to upscale by
545% 485% in the previous version only took 3 min in the new one.
Edit: corrected mistyped upscale percentage above.
-
Thanks for your report!
-
I just tried version 2.0 for the first time, and the speed increase is impressive. An image that took 19-20 minutes to upscale by 545% in the previous version only took 3 min in the new one.
Wow....that's not what I am experiencing. It is faster but no 3 min. processing on large files with high percentage upscale.
Can you tell me what size file you started with and was a jpg or tiff?
-
I can't speak for Arlen but I was getting about 2X speed (see my tests here (http://plugsandpixels.com/blog/released-today-topaz-a-i-gigapixel-2-0-5x-faster/)).
-
I was very surprised too, but it's not too far off from their claim of 5x faster. The 3 min was measured today with a timer; the 19-20 min was arrived at a week or two ago by glancing at the clock every so often, so it might be off by a minute or two.
The starting image was a 16 bit tiff, 5348 x 3395 px; upscaled to 25,920 x 16,454 px. The latter translates to 36" x 22.85" at 720 ppi, for test printing on an Epson printer.
-
I upgraded to a Nvidia 1070 gpu. Gigapixel processing that previously took minutes now takes seconds.
-
I upgraded to a Nvidia 1070 gpu. Gigapixel processing that previously took minutes now takes seconds.
Good to know, thanks.
Cheers,
Bart
-
Tested a 20 megapixel file from Sony RX100-4 with enlargement by 400%. Have Windows 10 on a 2-3 year old computer with fast SSD and 64 GB memory. Video card is 4 GB Nvidia. Time to process and write file was 4 minutes using version 2.0. I didn't test the earlier version. Invented detail on landscape shot was very convincing, particularly in showing fine detail of pine trees that were probably 400 yards away. I put a TIFF file crop (89 MB) on dropbox you can open in Photoshop to see the differences:
https://www.dropbox.com/s/o2xqgl45q31zd00/Gigapixel.tif?dl=0
There are 3 layers:
Photoshop Preserve Details (enlargement)
Gigapixel Reduce Noise/Blur None
Gigapixel Reduce Noise/Blur Moderate
My personal preference is half way between None and Moderate (set None layer to 50%) as it looks a little too smooth and plastic in the lake otherwise. Would be nice to have a Reduce Noise/Blur Mild. It is easier to see the differences that would probably appear in print by viewing at 50% in Photoshop. Original image attached with box showing crop in file on dropbox.
-
Thanks for sharing that file, John. Having the various options on layers is a good way to present it.
And I agree with your assessment that half way between None and Moderate works best in this case.
-
Tested a 20 megapixel file from Sony RX1004 with enlargement by 400%. Have Windows 10 on a 2-3 year old computer with fast SSD and 64 GB memory. Video card is 4 GB Nvidia. Time to process and write file was 4 minutes using version 2.0. I didn't test the earlier version. Invented detail on landscape shot was very convincing, particularly in showing fine detail of pine trees that were probably 400 yards away. I put a TIFF file crop (89 MB) on dropbox you can open in Photoshop to see the differences:
https://www.dropbox.com/s/o2xqgl45q31zd00/Gigapixel.tif?dl=0
There are 3 layers:
Photoshop Preserve Details (enlargement)
Gigapixel Reduce Noise/Blur None
Gigapixel Reduce Noise/Blur Moderate
My personal preference is half way between None and Moderate (set None layer to 50%) as it looks a little too smooth and plastic in the lake otherwise. Would be nice to have a Reduce Noise/Blur Mild. It is easier to see the differences that would probably appear in print by viewing at 50% in Photoshop. Original image attached with box showing crop in file on dropbox.
John, Thank you for sharing the comparison. The Preserve Details 2 gives significantly more detailed result. Since you have the original file, would be interesting to see how that compares to Topaz:)
-
John, Thank you for sharing the comparison. The Preserve Details 2 gives significantly more detailed result. Since you have the original file, would be interesting to see how that compares to Topaz:)
I modified the file on Dropbox to include a layer for Preserve Details 2. The Gigapixel result half way between Moderate and None still wins. Be sure to check at 100% and especially look at the top portion of the crop.
-
Thanks sportmaster and John for the feedback.
Maybe a Reduce Noise and Blur slider would be better than the three set options, I will suggest that.
-
Thank you John, I see a clear difference now. The texture of details generated by Gigapixel has much more natural and suiting appearance.
-
And another example.
- I started with a 1000 x 1000 pixel crop from a Capture One conversion of an image with very shallow Depth of Field, to which I added some FocusMagic Capture sharpening.
- I then downsampled the image to almost 1/6th, 167 x 167 pixels. This will introduce some resampling artifacts and a significant loss of actual resolution.
- I upsampled that back to 1000 x1000 pixels, with "Reduce Noise and Blur" set to None.
The result obviously has lower resolution than the original, but feel free to compare it to your own preferred recipe for upsampling. I'm confident about how well Topaz A.I. Gigapixel performed.
The first full-resolution crop also upscales very well to 600%, and AIG will add some interesting very high resolution down feather detail in the "None" setting (at the right side an below the eye). It also handles the difference between well-focused detail and blurred-detail in a pleasing manner for high quality printed output. Also look at how the sharpness of the iris is preserved.
Cheers,
Bart
-
This Topaz tool looks very promising. However after looking at the following review I was not so sure anymore:
https://youtu.be/LGWEyG4DUOM
It looks like Topaz is creating all kinds of very weird things that does not belong to the image in the first place.
What do you guys (and gals) think what's happening here?
Regards,
Jaap.
-
I'd seen this video before. Well done, but be aware, he's using version 1 from a month ago and really should do a revised version. Topaz software improves quickly and continually.
-
This Topaz tool looks very promising. However after looking at the following review I was not so sure anymore:
https://youtu.be/LGWEyG4DUOM
See my comparison earlier in this thread for a landscape image. I was skeptical, but the results speak for themselves. I purchased the program. PS I have no affiliation with Topaz.
-
My experience with Gigapixel is inconsistent. On some images the results are amazing. Fine, semi-linear details such as hair, feathers, mouths, etc. often render superbly. The details AI Gigapixel manufactures from nothing look realistic. Other times - I've seen it most often in eyes and textured fabrics - the added sharp details are unnatural and visually jarring. A enlargement made using traditional interpolation tools may be soft in comparison to Gigapixel but is visually preferable to a model with demonic eyes or clothing afflicted by leprosy.
-
My experience with Gigapixel is inconsistent. On some images the results are amazing. Fine, semi-linear details such as hair, feathers, mouths, etc. often render superbly. The details AI Gigapixel manufactures from nothing look realistic. Other times - I've seen it most often in eyes and textured fabrics - the added sharp details are unnatural and visually jarring. A enlargement made using traditional interpolation tools may be soft in comparison to Gigapixel but is visually preferable to a model with demonic eyes or clothing afflicted by leprosy.
If file size (<4 GB for TIFFs) is not an issue, one can layer 2 differently rescaled versions (both in AIG, or only one), and mask out the less successful regions.
When updates are issued, the A.I. may have been retrained or tuned to handle such features more gracefully.
Cheers,
Bart
-
For enlargements in the 2x - 3x range I am not seeing a lot of difference from PS for prints @ 20x30 when viewed at a normal viewing distance. Up close, maybe a slight edge to gigapixel.
Above 3x, starting to see this software pull ahead. Sharpening routines play a big part and I need to expirement further.
All in all I am satisfied with my purchase and look forward to further improvements.
-
I just had an opportunity to use Gigapixel to enlarge a small section of an 1877 map (https://bit.ly/2EkxDdm) showing a paper mill owned by the ancestors of one of my facebook friends.
The map only enlarged to a certain point online (as shown); Gigapixel did the rest.
-
For enlargements in the 2x - 3x range I am not seeing a lot of difference from PS for prints @ 20x30 when viewed at a normal viewing distance. Up close, maybe a slight edge to gigapixel.
Above 3x, starting to see this software pull ahead. Sharpening routines play a big part and I need to expirement further.
All in all I am satisfied with my purchase and look forward to further improvements.
With the uprezzing script developed 4 years ago here on lula, the best result was achieved when move the sharpering slider to 0, then resize, then do the sharpering. The outcome was always extraordinary.