The Big Picture (ACR Super Resolution)

[Chris Kern]

Where it works it works great, but at least in my tests with Z6 raw's in man made structures, there are a lot of color artifacts, it's like super resolution creates color aliasing and artifacts from clean sources.

As far as I have been able to determine, that's always a risk when you use this type of technology to enlarge files.  The process isn't simply interpolating extra pixels the way traditional enlargement techniques do; it's recreating the image in a larger format based on a machine-learning analysis of various attributes that it detects within the image.  If the software "discovers" an attribute that isn't really there from the perspective of a human viewer and then enhances it, you get what the viewer perceives as visual artifacts.  That risk can be reduced, as I understand it, by improving the training set of image components that is used in the machine-learning phase.

[kers]

After AI has done its job, you really should use your own CS and scrutinize/retouche the whole photo from top to bottom.

Then you are ready.

[Jonathan Cross]

Can anyone help, please?

I have Photoshop v22.2 installed on my late 2019 MacBook Air with 16GB memory and running Catalina.  I have ACR13.1 installed.  I understand that I need ACR13.2, but my Adobe update facility does not give me an option to download it.  As a result the only enhance option I have is Enhance Details. 

Is there away to get ACR13.2 so that I can use Super Resolution?

Thanks in advance for any advice.

Jonathan



   

[mcbroomf]

Usually when the Creative Cloud app doesn't show the update you need to log off CC (not just close it).  This forces you to log back on with your Adobe credentials when you re-open it and the updates will show up.

[Jonathan Cross]

Thanks Mike, but when I try to logout of the app, the sign out is greyed and I cannot!

Jonathan

[digitaldog]

Thanks Mike, but when I try to logout of the app, the sign out is greyed and I cannot!

Jonathan

Reboot the computer.

[mcbroomf]

Yes reboot.  If Sign Out is grey then it means you are not signed in which is why you're not seeing the updates.

[Jonathan Cross]

Thanks Andrew and Mike.  Rebooted Macbook and signed in to my Adobe account.  Creative Cloud app then stuttered into being, as it seemed what was there was an old version.  I got a message that it had been updated and would in future be updated.  It then spent a while updating Photoshop, Lightroom Classic and Camera RAW.  I now have ACR13.2, and have successfully done a super resolution.  It was of a Fuji 26MP RAF image that was some 56MB in size.  The super res DNG size is about 360MB (!) and the pixel count is the expected 4x the original.

Thanks for the advice.

Best wishes,

Jonathan



     

[Chris Kern]

I believe it will be rolled out to LR at some point. Thats good because my workflow is greatly slowed down opening an individual file in ACR directly. Then enhancing it creates a new DNG with the original filename and enhanced appended to it. I tried importing it into LR and while it gets aded top the catalogue LR is unable to view or output it.

Try the current (March) release of Lightroom (v. 10.2).  It now appears to parse the DNGs enlarged with Adobe Camera Raw without difficulty.  Other ACR image enhancements are propagated to LR via an xmp sidecar file.  Not seamless, but not overly cumbersome, either.

[Guillermo Luijk]

I asked a friend to test the algorithm against a synthetic image to check sharpness improvement on linear details (useful for texts and architectural images). The image was this:

http://guillermoluijk.com/misc/lozoya.png

And the result was quite surprising (I tried to fill some cells in red after the rescaling):



The improvement in detail is fine, but the algorithm also introduces some 'noise' (larger than one level in a 8-bit scale) as an undesired signal vs the expected values, surely as a result of its 'organic' AI nature.
Needless to say the image never went to any lossy format (JPEG), so all this variance comes from ACR's algorithm.

This could become irrelevant for real world photographs where read and photon noise will probably surpass this amount, but it's good to know.

Regards!

[Jonathan Cross]

Try the current (March) release of Lightroom (v. 10.2).  It now appears to parse the DNGs enlarged with Adobe Camera Raw without difficulty.  Other ACR image enhancements are propagated to LR via an xmp sidecar file.  Not seamless, but not overly cumbersome, either.

+1

Jonathan

[Benny Profane]

I asked a friend to test the algorithm against a synthetic image to check sharpness improvement on linear details (useful for texts and architectural images). The image was this:

http://guillermoluijk.com/misc/lozoya.png

And the result was quite surprising (I tried to fill some cells in red after the rescaling):



The improvement in detail is fine, but the algorithm also introduces some 'noise' (larger than one level in a 8-bit scale) as an undesired signal vs the expected values, surely as a result of its 'organic' AI nature.
Needless to say the image never went to any lossy format (JPEG), so all this variance comes from ACR's algorithm.

This could become irrelevant for real world photographs where read and photon noise will probably surpass this amount, but it's good to know.

Regards!

Hey, nothing's free in life.

[Guillermo Luijk]

Two animated GIF's. First shows the original synthetic image (800%) vs superresolution (400%):

http://guillermoluijk.com/misc/superresolucion.gif

I feel the rescaling enhances detail and sharpness. Another good thing is that it manages to recreate barely visible cells surrounded by close edges in the original image as correct polygonal shapes.
On the downside, noise is already visible. Let's enhance it with a curve:

http://guillermoluijk.com/misc/superresolucioneditada.gif

Noise spreads to areas far from the black edges, where the original image was pure white. Moreover it is coloured noise, while the original image was pure grayscale (Adobe could have taken care of this in case someone rescales monochrome images).

Regards

[SmartSolutions]

Thanks for posting this topic it will be great help

[Jonathan Cross]

Right or wrong?

If I have an iphone image of 12MP and up res it using super res in Photoshop or Gigapixel, I can have a 48MP image.  If I have an image of the same scene taken with, say, a Sony a7r4 and then cropped to 48MP and the same coverage as the other image, which will have more detail?  I think it will be the Sony - right or wrong?

Jonathan

[Chris Kern]

If I have an iphone image of 12MP and up res it using super res in Photoshop or Gigapixel, I can have a 48MP image.  If I have an image of the same scene taken with, say, a Sony a7r4 and then cropped to 48MP and the same coverage as the other image, which will have more detail?

Even assuming you started with a pristine (raw or at least uncompressed) file from the iPhone, you would be comparing a recreated image with an actual one.  The AI enlarging tools based on machine-learning analyze the components of an image and reconstitute it in a larger format based on the visual elements they have recognized in the source file.  Assuming they do a good job and don't introduce any objectionable artifacts—and in my experience, they often do, but not always—they may be more effective on a given image than enlarging it by interpolating more pixels (e.g., in Photoshop).  But I can't imagine a situation when you wouldn't be better off working from the same amount of data collected directly by a camera's light sensor.  The value of these tools is to give you more apparent resolution than is available from the camera on those occasions when you really need it.

[Chris Kern]

The AI enlarging tools based on machine-learning analyze the components of an image and reconstitute it in a larger format based on the visual elements they have recognized in the source file.

Apropos of which, it's important to select an appropriate "training set" of images which will be fed to the neural network during the recognition phase in order to achieve the desired result.  Train the software on images seen through a window covered with raindrops and it will find them in your photo whether they exist there or not.  Attached: (1) a source image of the George Washington Bridge; (2) a target image produced by a convolutional neural network trained to find raindrops.

[John Hollenberg]

Apropos of which, it's important to select an appropriate "training set" of images which will be fed to the neural network during the recognition phase in order to achieve the desired result.  Train the software on images seen through a window covered with raindrops and it will find them in your photo whether they exist there or not.  Attached: (1) a source image of the George Washington Bridge; (2) a target image produced by a convolutional neural network trained to find raindrops.

I rather like the "bridge in the rain" result.  It appears that the neural network has created a new reality.

[Chris Kern]

I rather like the "bridge in the rain" result.  It appears that the neural network has created a new reality.

This is an example of what I think of as using a convolutional neural network backwards.  Instead of having it identify raindrops in an image, use it to recognize raindrop-like visual elements in the source file and make them look more like raindrops in the destination file.  Instead of trying to recognize a painting by van Gogh, inject van Gogh-like attributes into the image.

I experimented with this technique on a number of photographs a while back.  Some produced interesting results, others didn't.  You can see some samples here.

[kers]

One would like to use super resolution to see more detail in a photo.
indeed at 100% pixel size many details seem improved, yet it goes along with some side effects that are larger than just details

Here two problems i encounter quickly when using the super resolution introduced in 22.3 photoshop.

two versions- 100% pixel size
org- means normal developed and added some basic sharpening in photoshop
better- means superresolution- no sharpening added.

example1 
The arrow points at areas with introduced artefacts.
example2
the arrow points at uneven areas with blur/detail that used to be even.

if i use the image smaller on the web, I can see no improvement in detail, but i can see the introduced artefacts.
If i use the image at 300dpi in print I cannot see any difference, but the introduced artefacts.
If i use the image at 100dpi in print I can see the better detail of SR, but also the artefacts.

ergo: in only a small number of cases the SR has an advantage, but the images should be scrutinized to get rid of all the false artefacts.
I am sure that in future versions of photoshop ( or Topaz resize) things will improve.