Luminous Landscape Forum
Raw & Post Processing, Printing => Digital Image Processing => Topic started by: John Hollenberg on September 29, 2019, 05:27:16 pm
-
On1 Resize: https://blog.kasson.com/nikon-z6-7/megapixels-smart-resizing-on1/
Topaz Gigapixel AI: https://blog.kasson.com/nikon-z6-7/megapixels-smart-resizing-and-printing-topaz-gigapixel-ai/
Alien Skin Blowup: https://blog.kasson.com/nikon-z6-7/megapixels-smart-resizing-and-printing-alien-skin-blow-up/
Photozoom Pro 8: https://blog.kasson.com/nikon-z6-7/megapixels-smart-resizing-and-printing-photozoom-pro-8/
My conclusion: Gigapixel AI is the one to own (I do)
-
For me personally I still can’t accept the artifacts that Gigapixel AI provides, very unnatural. I am positive that Topaz will get there in the end and since the money has been spent I’ll be patient ;)
In the mean time I need to settle for second best. For small upscale factors I’ll use PS and C1 but for larger upscales I’m leaning towards Photozoom Pro8.
Same is applicable w.r.t. Denoise AI and Sharpen AI. NeatImage and FocusMagic give me less unwanted artifacts and a more natural, ‘real’ image.
So far my investment in the Topaz AI bundle hasn’t been successful. Nevertheless I find the technology behind it promising for the (near) future.
Regards,
Jaap.
-
For me personally I still can’t accept the artifacts that Gigapixel AI provides, very unnatural. I am positive that Topaz will get there in the end and since the money has been spent I’ll be patient ;)
Hi Jaap,
I'm not sure what kind of artifacts you are talking about, but I have noticed some differences in results between CPU processed results and GPU processed ones. The CPU processing, has recently been sped up a lot for Intel CPUs, by the use of the optional Intel OpenVINO. Check out the Advanced Preference settings. Even my older generation Intel CPU processes images more than 2x faster (or less slowly ;)).
In the mean time I need to settle for second best. For small upscale factors I’ll use PS and C1 but for larger upscales I’m leaning towards Photozoom Pro8.
Like Jim Kasson, overall I prefer the Gigapixel results. Do take into account the actual size the image detail will have in the output. It will look much better than when pixel peeping the zoomed-in pixels at close proximity on one's monitor display.
Cheers,
Bart
-
Hi Bart,
Most of the artifacts I see are around high contrast areas, i.e. tree branches and leaves against a bright sky, resulting in various kinds of artifacts. Also around edges of buildings and lamp posts, resulting in a kind of ghostly images many pixels away from the original edges. To me this is so unnatural that it is not usable.
For completeness I need to mention that I’m taking special care not to oversharpen my images.
I only apply GPU processing with my Geforce 1070Ti card. I have zero speed issues.
I fully agree with you that on Jim’s images Gigapixel AI is preferable over the rest.
Regards,
Jaap.
-
Hi Bart,
Most of the artifacts I see are around high contrast areas, i.e. tree branches and leaves against a bright sky, resulting in various kinds of artifacts. Also around edges of buildings and lamp posts, resulting in a kind of ghostly images many pixels away from the original edges. To me this is so unnatural that it is not usable.
I understand.
Try no sharpening. I'm a Capture One user, so that still produces high resolution conversions (potentially with added 'structure'), and it reduces the risk of sharpening noise. Upscale that with Gigapixel (which still adds resolution and offers to increase sharpness), and optionally sharpen that result. When the source image has low noise, the noise reduction in Gigapixel can be set to very low levels, which brings out more detail and structure where it's supposed to be.
Cheers,
Bart
-
Here are some GPU vs CPU results with Win 7 x64 i7 4700MQ with Nvidia GeForce GT 740M and/or Intel(R) HD Graphics 4600 running on an old Toshiba laptop.
Gigapixel seems to "see" the Intel(R) HD Graphics 4600 in its Graphics Info dialog.
Gigapixels GPU processing creates artifacts as exemplified in the attached images. The images were upsized in Gigapixel approximately 2.83x. The crops shown here are 200% pixel peeps. You wouldn't see the artifacts in a 300dpi print, but they are still disconcerting.
-
In my test between 6 software: Adobe Photoshop (CC 2018), BenVista PhotoZoom Pro 7, ON1 Resize 2018, Topaz Gigapixel AI, Alien Skin Software Blow Up 3, Qimage Ultimate 2019.119, Gigapixel AI absolutely gets best performance.
I tested 22 interpolation methods and 3 upscale ratios on 4 images, that's 264 images. full test article here: part 1 (https://mfalab.pixnet.net/blog/post/32786686) | part 2 (https://mfalab.pixnet.net/blog/post/32796406)
Conversely, for downsize application, bicubic is good enough.
-
This is of recent interest to me, so thanks for all the hard work some of you have put in.
Just thinking out loud, I wonder if there is a clear winner when one processes files from different camera brand raw files. In particular the Fuji X-Tran sensors seem to get significantly different looks from different Raw processing programs, and even different version of the same converter.
I wonder If that is also something to be factored in, with these resining programs?
-
IMHO many RAW processing products do a mediocre or merely an acceptable job demosaicing Fuji X-Trans files. The best I’ve found for my X-Trans files is Iridient X-Transformer. Normally I use it on my files that have a need to show details. Others may feel differently.
-
This is of recent interest to me, so thanks for all the hard work some of you have put in.
Just thinking out loud, I wonder if there is a clear winner when one processes files from different camera brand raw files. In particular the Fuji X-Tran sensors seem to get significantly different looks from different Raw processing programs, and even different version of the same converter.
I wonder If that is also something to be factored in, with these resining programs?
Hi Pavel,
Topaz Gigapixel AI will work on whatever it gets presented, so the better the input quality, the better the output quality will be. I don't have personal experience with X-Trans files, but FWIW I've heard positive reactions about Capture One conversions.
Cheers,
Bart
-
Topaz Gigapixel AI will work on whatever it gets presented, so the better the input quality, the better the output quality will be. I don't have personal experience with X-Trans files, but FWIW I've heard positive reactions about Capture One conversions.
I can confirm this. But be very careful with (pre)sharpening. I get best results applying no sharpening at all in C1, otherwise Gigapixel creates nasty artifacts.
Regards,
Jaap.
-
I can confirm this. But be very careful with (pre)sharpening. I get best results applying no sharpening at all in C1, otherwise Gigapixel creates nasty artifacts.
Regards,
Jaap.
Likewise.seems to work well with Xtrans (no presharpening)
-
Interesting discussion, just so I can keep my understanding in line a simple question. Does GigaPixel work with raw file data?
Or does one have to use their raw processor of choice to create a tiff to send to Gigapixel?
-
In my test between 6 software: Adobe Photoshop (CC 2018), BenVista PhotoZoom Pro 7, ON1 Resize 2018, Topaz Gigapixel AI, Alien Skin Software Blow Up 3, Qimage Ultimate 2019.119, Gigapixel AI absolutely gets best performance.
I tested 22 interpolation methods and 3 upscale ratios on 4 images, that's 264 images. full test article here: part 1 (https://mfalab.pixnet.net/blog/post/32786686) | part 2 (https://mfalab.pixnet.net/blog/post/32796406)
Conversely, for downsize application, bicubic is good enough.
Interesting research. However it is not in-line with my findings. Too often funny artefacts with gigapixel. Also the image appears not real anymore. (Do not know how to word how it appears to me). So i do not use it anymore. I still use the tool developed in Lula in 2014, or Photoshop.
-
Interesting discussion, just so I can keep my understanding in line a simple question. Does GigaPixel work with raw file data?
Or does one have to use their raw processor of choice to create a tiff to send to Gigapixel?
Hi Denis,
Yes, Gigapixel AI does read Raw files, but it doesn't offer color balancing or lens corrections. So for best quality, I'd feed it a TIFF conversion with minimal (if any) sharpening.
Cheers,
Bart
-
Interesting research. However it is not in-line with my findings. Too often funny artefacts with gigapixel. Also the image appears not real anymore. (Do not know how to word how it appears to me). So i do not use it anymore. I still use the tool developed in Lula in 2014, or Photoshop.
Gigapixel has strange artifact when it made a bad guess of object in image. Mix photoshop preserve details 2 image is better way in practice. I will do frequency separation on gigapixel image, bottom layer is preserve details 2 image. Now I have 3 layers, gigapixel low frequency layer is masked some "too clean" area to show a little noise from preserve details layer. And Adjust opacity of high frequency layer from gigapixel to optimize more nature look. I can apply mask to block some artificial area if necessary.
Gigapixel is not a perfect tool yet. But AI computing is the future way.
-
I am playing around with various AI labeled software options. Turns out the Topaz plugins I bought ages ago and wasn’t using entitled me to various upgrades such as Sharpen and denoise. I decided to buy AI gigapixel at the same time as I upgraded the free stuff and paid for AI masking.
I am actually astonished at some of the results I am getting. Not sure all the software is ready for prime time but I agree that it is the future. Sort of a computational set of tools for serious photographers. It has actually caused me to reassess some camera gear upgrade plans.
-
I've downloaded Gigapixel AI and have been experimenting with it. One trivial question. I do most of my work and all my printing with LR. When you export a RAW file converted to a TIF for Gigapixel, does it matter what color space it's in? I will be exporting back into LR following uprez. I ran a couple yesterday and it seems to go fairly smoothly. I have a pretty fast i7 CPU but the graphics card has only 2GB RAM and this might make me upgrade.
-
I've downloaded Gigapixel AI and have been experimenting with it. One trivial question. I do most of my work and all my printing with LR. When you export a RAW file converted to a TIF for Gigapixel, does it matter what color space it's in?
Hi Alan,
Gigapixel AI is supposed to recognize the (e.g. TIFF) embedded Colorspace, and it offers to convert and tag the resulting image in a few commonly used colorspaces, or keep the original colorspace.
Cheers,
Bart
-
Hi Alan,
Gigapixel AI is supposed to recognize the (e.g. TIFF) embedded Colorspace, and it offers to convert and tag the resulting image in a few commonly used colorspaces, or keep the original colorspace.
Cheers,
Bart
Yes, that part was apparent. I guess since LR uses ProPhoto as the colorspace, I should export in that and there would not be any issues upon importing the up-rez file.
-
I have a pretty fast i7 CPU but the graphics card has only 2GB RAM and this might make me upgrade.
At this time Gigapixel does not make full use of GPU memory (yet?). Upscaling to maximum pixel resolution of 32000 px (8 bit) / 22000 px (16 bit) on one side (3:2) uses 58.8% of my 8 gb GPU memory when GP is set to "High" memory consumption. So 4 to 6 gb should be sufficient for the time being, with 8 gb being more future-proof.
-
In my test between 6 software: Adobe Photoshop (CC 2018), BenVista PhotoZoom Pro 7, ON1 Resize 2018, Topaz Gigapixel AI, Alien Skin Software Blow Up 3, Qimage Ultimate 2019.119, Gigapixel AI absolutely gets best performance.
I tested 22 interpolation methods and 3 upscale ratios on 4 images, that's 264 images. full test article here: part 1 (https://mfalab.pixnet.net/blog/post/32786686) | part 2 (https://mfalab.pixnet.net/blog/post/32796406)
Conversely, for downsize application, bicubic is good enough.
Hello MfAlab,
Thank you for the comparison,
There are two things i do not like about Gigapixel AI;
1 it changes local contrast- more than the other approaches.
2 it produces false detail and simplifies the image; making it look like a vectorimage
In my own tests these two problems were visible on print as well- and that makes it serious.
I am in favour of photoshop 2.0...
Here some images of you to stress my point- and thank you for providing them.
for local contrast i introduced a gaussian blur to make it a bit more visual.
-
I guess since LR uses ProPhoto as the colorspace....
Not really. But we can go there.
I should export in that and there would not be any issues upon importing the up-rez file.
You should export in the color space you desire for the task at hand, that may very well be ProPhoto RGB from the internal color space LR uses for processing which isn't ProPhoto RGB (close but no cigar).
If you really, really wanted to export using LR's internal processing color space, that is possible.
-
Hello MfAlab,
Thank you for the comparison,
There are two things i do not like about Gigapixel AI;
1 it changes local contrast- more than the other approaches.
Hi Pieter, there have been several updates on the AI models. Maybe it has changed in more recent versions?
2 it produces false detail and simplifies the image; making it look like a vectorimage
Try reducing the amount of noise reduction in GAI, and Gigapixel AI will create lots of subtle detail. This may not result in smooth skies/gradients, but that's a trade-off (or one could do 2 renderings and combine them in a masked Layer blending). Also try CPU processing and compare it to GPU processing. CPU processing can result in more detail.
Cheers,
Bart
-
Not really. But we can go there. You should export in the color space you desire for the task at hand, that may very well be ProPhoto RGB from the internal color space LR uses for processing which isn't ProPhoto RGB (close but no cigar).
If you really, really wanted to export using LR's internal processing color space, that is possible.
All I want to do is export it in a form that won't screw things up down the line. I'm just in the testing mode right now with this software and wanted to see how it performs before spending the $100 for a license. After doing the up-rez, I will import back into LR to see 1) if anything else needs to be done on the image, and 2) to print it. I would be creating an image that is set for a specific print size and ppi. I'm on a Canon Pro-1000 (17 inch printer) so any large images would be set for 300 ppi. Most of the Z6 images are 24 megapixel so there would no large increases to get to the desired image size since I would likely have either 15 or 16 inch to get to where I want. I have a lot of older images from my Nikon D300 which are smaller in pixel size and might benefit more. I'm just curious how much better this is than the LR up-rez I have used in the past. I pretty much follow Jeff Schewe's approach to up-rez of images.
-
Hello MfAlab,
Thank you for the comparison,
There are two things i do not like about Gigapixel AI;
1 it changes local contrast- more than the other approaches.
2 it produces false detail and simplifies the image; making it look like a vectorimage
In my own tests these two problems were visible on print as well- and that makes it serious.
I am in favour of photoshop 2.0...
Here some images of you to stress my point- and thank you for providing them.
for local contrast i introduced a gaussian blur to make it a bit more visual.
I set every setting bar at the middle (or default) for every software to do a "preset test". But it's too much for AI gigapixel. The detail and contrast problem can be reduce by tuning the settings. AI gigapixel updates continually. I was tested on 2.X version. Maybe newer version perform better.