Luminous Landscape Forum
Raw & Post Processing, Printing => Digital Image Processing => Topic started by: bobfriedman on April 20, 2019, 05:07:00 pm
-
just getting lucky today..
Nikon D800M,Carl Zeiss Otus 1.4/55
1/60s f/4.5 at 55.0mm iso5600
(https://pbase.com/bobfriedman/image/169084061/original.jpg)
AI Sharpen using Stabliize with Denoise at 75%
(https://pbase.com/bobfriedman/image/169084079/original.jpg)
-
just getting lucky today..
Nikon D800M,Carl Zeiss Otus 1.4/55
1/60s f/4.5 at 55.0mm iso5600
Hi Bob,
Feeding a Monochrome image (from an 800M if that's not a typo) to an application that is trained for color image input is not the best way to use that application.
I also wonder about the ISO setting as such, you'll probably get better noise performance with an ISO 1600 setting, and underexposing that if you need the short exposure time. Then in postprocessing compensate for the underexposure and know that you'll probably have full detail in the highlights.
Cheers,
Bart
-
if you remember I got this with a color image as well..
https://forum.luminous-landscape.com/index.php?topic=129798.0
note: D800M is my own nomenclature to refer to a D800 that has been sensor modified to monochrome.
-
I also wonder about the ISO setting as such, you'll probably get better noise performance with an ISO 1600 setting, and underexposing that if you need the short exposure time.
Actually the Museum of Fine Art was so dark I had really no choice. Images were in fact underexposed. Not my first choice, believe me.
-
I have been assuming the patterns are artifacts of a divide, compute, and reassemble process that creates an array of subsets which may, or may not, be effectively stitched back together to appear seamless.
I similar pattern artifacts while demonstrating GigaPixel.
Topaz explained that they were aware of the issue, and suggested the artifacts are the result of a compromise that makes the computation times practical.
-
well, I decided to process the NEF file with Nikon Capture NX-D rather than ACR in Photoshop..
I employed a standalone AI Sharpen with Stabilize and Denoise at 80% at arrived at the image below free of banding.
(https://pbase.com/bobfriedman/image/169087844/original.jpg)
-
well, I decided to process the NEF file with Nikon Capture NX-D rather than ACR in Photoshop..
I employed a standalone AI Sharpen with Stabilize and Denoise at 80% at arrived at the image below free of banding.
Hi Bob,
Thanks for persevering. Who'd have thunk that, different Raw converter, such a different result ... Same settings in SAI?
I bet that the folks at Topaz would love to have both original examples, even though they are Achromatic. I know they are looking into the "Grid" issue that occasionally pops up on some images, and this might steer them in the direction of a solution. Their forum is closed for a planned upgrade right now, but Service Requests can always be opened via their Help Center when logged in.
Cheers,
Bart
-
This might be another case where Topaz Sharpen AI is exacerbating subtle flaws in the original file. I can see hints of banding/blocking in the ACR (unsharpened) image. Perhaps the NX-D version does not have those artifacts.
-
This might be another case where Topaz Sharpen AI is exacerbating subtle flaws in the original file. I can see hints of banding/blocking in the ACR (unsharpened) image. Perhaps the NX-D version does not have those artifacts.
this happened with AI Denoise as well... so you wouldn't think that would amplify errors but he opposite.
-
this happened with AI Denoise as well... so you wouldn't think that would amplify errors but he opposite.
Hi Bob and Julian,
Actually, this (potential) Raw converter dependency is a very fascinating observation that will take further investigation to validate. If so, it might suggest that in using the specific images used for training (the training set), the various models were derived from under-represented common ACR conversions. Then feeding ACR conversions, the supposed grid structure would look to an AI system as genuine detail to preserve and enhance, in whatever mode it was operating.
Again speculating, if so, it could lead to improvements in Raw conversion and/or of the "training set". Either way, it could lead to better tools.
Cheers,
Bart
-
If you look carefully you can see the banding is also buried deep in tonal range of the first image and the third image.
I down loaded the third and use Clarity to emphasize the grid pattern. I will not post a copy of your photo, but you may wish to try it to see what is buried in there.
I think my previously voiced assumption was probably a mistake.
-
There's now also a thread on Topaz Sharpen AI on GetDPI on this: Topaz artefacts (https://www.getdpi.com/forum/image-processing/66058-topaz-artefacts.html#post786067)
Here's my post with some test results:
I exported an image from about 2000 pixels from Lightroom, it gave very similar artifacts in Sharpen AI along the blue sky top border.
so it seems C1 is not the culprit. I thought it was very unlikely that it would, but now we're sure.
Edit 2: Here's my two examples:
Exported from LR, without any sharpening (slider at zero in the converter, none in the export dialog)
(https://photos.smugmug.com/photos/i-zvHwkjK/0/e5b9bd39/O/i-zvHwkjK.jpg)
Above image, sharpened in Topaz AI (default parameters) ==> Artifacts galore along the top border
(https://photos.smugmug.com/photos/i-9czLsHh/0/87371bc5/O/i-9czLsHh.jpg)
And added just for fun, first image, sharpened with Topaz Detail (minor small and middle details and .10 deblur) ==> No artifacts and slightly different result from sharpen AI on the statue, can't really decide which is better, they're just a tiny bit different.
(https://photos.smugmug.com/photos/i-TddcKcK/0/13145018/O/i-TddcKcK.jpg)
Further it was suggested in the thread that the raw converter can have an effect but the thread shows there is little difference between Lightroom and C1, and I can't imagine Topaz not training their AI models on these two quite popular and widely used raw converters.
It was further tested and the artifacts do not appear when exporting the file without downsizing and then running through Sharpen AI
However exporting a downsized tiff (8 bit and 16 bit) gave the same artifacts as running sharpen AI on a jpeg.
So it seems downsizing is the culprit leading to the artifacts in solid colour areas.
I have reported this issue to Topaz this is what I got back:
I just tested your image on my computer and I can reproduce your issue. I will discuss with the leader of our R&D team to see if we can figure out the reason of this issue.
My suggestion for now is please try to use Sharpen AI before Lightroom adjustments. Our AI model sometimes can be sensitive and typically can do the best job on the very original image.
I find the advice kind of lame, since final sharpening for output size is usually done as the last step and at output resolution (at least that's how I get the best results, YMMV) and not as a first step before making any adjustments. I find capture sharpening in Lightroom (and C1) more than adequate and don't need Topaz sharpen AI for that step.
I'd be interested in your experiences (similar or otherwise) or some suggestions how to avoid this issue on downsized raw converter exports.
-
If it gives artifacts, why use it? The whole point is to make the image better, not worse.
-
If it gives artifacts, why use it? The whole point is to make the image better, not worse.
I agree... it is hard for me to accept that I need to be very careful using this tool.
-
If it gives artifacts, why use it? The whole point is to make the image better, not worse.
That's the wrong question in my mind, for me the real question is: "can Topaz fix it so it doesn't give artifacts and make it a useful tool"
-
That's the wrong question in my mind, for me the real question is: "can Topaz fix it so it doesn't give artifacts and make it a useful tool"
And there is hope ;)
Hi Pieter,
Thanks for your reply!
I just checked our current models (which are NOT yet released) for Sharpen AI. I believe the new models can do a better job on your images. In fact, I just tried your image and the annoying artifacts are gone!
There will be a release for Sharpen AI in early May. We will continue improving our models. Thanks for your patience!
Sincerely,
Topaz Labs Support
Let's wait and see what the update in May brings.
-
And there is hope ;)
Welcome back! Where you've been the whole year? :)
-
Welcome back! Where you've been the whole year? :)
Thanks Slobodan, I took a sabattical from Luminous Landscape and concentrated on photography and other ventures.
Kind of grew tired of some discussions with kindergarten pupils imposing as wise old men ;).
Still am btw, but the notion of Bart van der Wolf trying to sort through some of these Topaz technical issues made me come back for that.
-
Thanks Slobodan, I took a sabattical from Luminous Landscape and concentrated on photography and other ventures.
Kind of grew tired of some discussions with kindergarten pupils imposing as wise old men ;).
Still am btw, but the notion of Bart van der Wolf trying to sort through some of these Topaz technical issues made me come back for that.
Well put! The nuggets are worth hunting out...but at times the tailings are unecessarily large (and mostly unecessary fullstop)
-
There's now also a thread on Topaz Sharpen AI on GetDPI on this: Topaz artefacts (https://www.getdpi.com/forum/image-processing/66058-topaz-artefacts.html#post786067)
Here's my post with some test results:
Further it was suggested in the thread that the raw converter can have an effect but the thread shows there is little difference between Lightroom and C1, and I can't imagine Topaz not training their AI models on these two quite popular and widely used raw converters.
It was further tested and the artifacts do not appear when exporting the file without downsizing and then running through Sharpen AI
However exporting a downsized tiff (8 bit and 16 bit) gave the same artifacts as running sharpen AI on a jpeg.
So it seems downsizing is the culprit leading to the artifacts in solid colour areas.
I have reported this issue to Topaz this is what I got back:
I find the advice kind of lame, since final sharpening for output size is usually done as the last step and at output resolution (at least that's how I get the best results, YMMV) and not as a first step before making any adjustments. I find capture sharpening in Lightroom (and C1) more than adequate and don't need Topaz sharpen AI for that step.
I'd be interested in your experiences (similar or otherwise) or some suggestions how to avoid this issue on downsized raw converter exports.
Pegelli, did you find a solution yet to those artifacts? I'm seeing the exact same sort of thing for some images that go through Gigipixel AI, rather than Sharpen AI. The pattern that you showed repeats at regular intervals along the top border for some (but not all) images with smooth blue skies. Hopefully if Topaz figures out how to fix it in one of the applications, it will work for both.
-
I think the product is not mature and it is not very helpful that you have too scrutinize the whole image to see if something is messed up after using Sharpen AI.
At most it can save an already bad image in some occasions. That can be important of course, but for general purpose i would not like to use it.
With todays lenses and sensors, the sharpen- as well as the noise-problems are not so relevant anymore.
Most noise that you see 100% on screen almost completely disappears in print. On screen the 100% images is not used in 99% of the occasions, usually a 25% image and then noise is hardly an issue.
The artifacts and uneveness produced here i find much worse than a regular noise pattern.
-
Kind of grew tired of some discussions with kindergarten pupils imposing as wise old men ;).
Hmmm.. IMHO this level of arrogance is off topic.
-
I think the product is not mature and it is not very helpful that you have too scrutinize the whole image to see if something is messed up after using Sharpen AI.
Quite agree. I think this applies to the entire AI suite.
-
I find all three of the AI programs that I have (disregarding JPEG to RAW AI, which I haven't tried) useful enough to be glad that I own them. But I will be glad when the bugs are worked out. Actually the bug that currently bothers me the most, and should be the easiest to fix, is that Sharpen AI totally messes up a PS Smart Object workflow. If you use any Topaz Studio adjustment/application on an SO, it comes back labeled as Sharpen AI. And if you click on that smart layer to edit it, it opens up as Sharpen AI, too. Even if you haven't used Sharpen AI at all. So the ability to edit Topaz Studio applications on a smart object is lost. This bug has been reported by many people for months. I think the popularity of these programs has overwhelmed Topaz, and they need some more staff.
-
I think the popularity of these programs has overwhelmed Topaz, and they need some more staff.
Or perhaps they should wait until the bugs are worked out before releasing a new version. I have held off purchasing any of their products because of all the bugs reported here and elsewhere.
-
I agree. I think it is however a pretty big achievement with these Topaz AI programs to perform better than the currently best in class programs such as Neat Image and Focus Magic. This is unfortunately on a case by case basis and the results are currently not reliable, certainly not in a batch workflow.
Recently I clearly saw good incremental achievements made, especially on AI Gigapixel, and as far as I can see the future looks promising but indeed the bugs need to be sorted out. Topaz needs to prove themselves that their audience is not the hobby kind of snappers but the pro photographers as well, in need for quality, reliability and reproducibility.
Regards,
Jaap.
-
There's now also a thread on Topaz Sharpen AI on GetDPI on this: Topaz artefacts (https://www.getdpi.com/forum/image-processing/66058-topaz-artefacts.html#post786067)
Here's my post with some test results:
Further it was suggested in the thread that the raw converter can have an effect but the thread shows there is little difference between Lightroom and C1, and I can't imagine Topaz not training their AI models on these two quite popular and widely used raw converters.
It was further tested and the artifacts do not appear when exporting the file without downsizing and then running through Sharpen AI
However exporting a downsized tiff (8 bit and 16 bit) gave the same artifacts as running sharpen AI on a jpeg.
So it seems downsizing is the culprit leading to the artifacts in solid colour areas.
I have reported this issue to Topaz this is what I got back:
I find the advice kind of lame, since final sharpening for output size is usually done as the last step and at output resolution (at least that's how I get the best results, YMMV) and not as a first step before making any adjustments. I find capture sharpening in Lightroom (and C1) more than adequate and don't need Topaz sharpen AI for that step.
I'd be interested in your experiences (similar or otherwise) or some suggestions how to avoid this issue on downsized raw converter exports.
It seems like the artifacts in your test image appear as a result of the algorithm struggling with the uniform area of sky next to the edge of the image. I took your top image (LR output) into PS and extended the canvas from 1333 pixels tall to 1600 pixels tall, with the canvas extension applied to the top of the canvas. I duplicated the original image and flipped it vertically, mirroring the sky pixels along the top edge and giving the extended border a healthy 250+ pixels of mirrored extension. I ran Sharpen AI at default (Sharpen mode) on both the original, where I reproduced the artifacts that you demonstrate, and on the extended, mirrored version. The artifacts appear on the extended, mirrored version, but at the edge of the image again, leaving a clean image at the original edge of the image, where the mirrored pixels are. The artifacts areas appear regularly spaced along the boundary of the image and are arranged in similarly appearing clusters.
These artifacts are analogous to edge artifacts from FFT filtering or naive deconvolution or tone mapping a 360° x 180° equirectangular panorama without repeating edge pixels or having an edge-aware tone mapping algorithm.
See attached: red arrows indicate the mirror axis for the repeated edge pixels; yellow arrows indicate artifact clusters.
Kirk
-
I also tried an experiment with your image where I used a vertical pixel offset (295 pixels downward) to make the cement wall wrap around the bottom of the image and appear at the top of the image. This offset aligned the original top of the wall with the top edge of the image boundary, so that the image appears to have the cement wall hanging from the top of the image. I ran Sharpen AI on the image, default Sharpen values, and no artifact appeared along the top edge of the offset image. Upon applying a negative offset to reposition the image to its original appearance, the image is artifact-free, without having to mirror the broad, uniform sky area pixels to achieve the same output.
So, it would appear that the top edge of the image may be prone to artifact if it contains areas of broad, uniform tone/color, whereas it will be okay if it contains detail.
A kludge to experiment with and verify might be that if you are getting artifacts in images with broad, uniform areas of color or tone at the top of your images, offset the image so that detailed areas of the image are positioned along the top of the image and then run SharpenAI, then apply negative offset to return the image to its original position.
Kirk
-
Thanks for those experiments, Kirk. Your results are consistent with what I've been seeing with not only Sharpen AI, but also Gigapixel AI.
-
The recent program update (about a week ago) fixed the issue for me.
I spotted no more artifacts in downsized LR or C1 exports when applying Sharpen AI.
Also the Denoise AI seems to work a lot better, mainly less artifacts in smooth areas close to highly detailed ones.
See here for the latest discussion on GetDPI on the Topaz AI programs. (https://www.getdpi.com/forum/image-processing/66058-topaz-artefacts.html#post788779)