Upscaling with LRC

[AL1M]

LRC 14.3.1 CR 17.3.1

Despite my extensive research, I can't find a clear, single answer online. I can't figure out how LRC works, and therefore if there's a less quality-degrading method.

I know of three different ways to modify your composition while maintaining the same number of pixels as the original photo, knowing that cropping a photo reduces its pixel size, which sometimes poses problems later on.

One way is to crop and then export at the original image size.

A second way is to use the scale slider in the transform tool to "Zoom" into the image to the desired composition, thus avoiding reducing the original pixel size.

A third way is to first use super resolution. This will give you twice the width and twice the height. Then crop to the desired composition and finally export at the original image size.

And of course, ditch Lightroom Classic and switch to Photoshop for cropping and resizing.

Isn't the difference between these three methods obvious? Does this mean that all three solutions are equivalent in terms of processing by the software?

Thank you for your insight.

[Per Berntsen]

For a small amount of upscaling you can just export at the desired dimensions.

The upscaling algorithm does a very good job, up to a point.

For larger amounts of upscaling I would use Super resolution, and then export at the desired dimensions.

There is no need for Photoshop here, unless you need a very precise crop, which is difficult to achieve in LrC.

Experiment with these two methods, and see what works best for you.

I have never used Transform for upscaling, but I suspect that it's only designed for small corrections.

But you can try it, and compare with the two other methods.

[AL1M]

Thank you. transformation allows the increase up to 1.5. I did the various tests. There does not seem to be any real difference. It is possibly necessary to redo a little sharpening except by going through super resolution. I would however like to know a little more about the process and the method of treatment actually carried out and of course on the real order of the different successive treatments.

[Conrad_C · Accepted solution]

This is how I understand it:

Crop and export at original size. I understand this to mean you didn’t apply changes in the Resize To Fit options. If that’s the case then the export should be an exact export of the original pixels, so it should look the same as the original but without the pixels you cropped out. So the width and height in pixels decrease compared to the original. If you did enter something in Resize to Fit, then scaling/resampling will be applied of course. The question is, what kind is it? The handy book “Adobe Lightroom Classic: The Missing FAQ” says Resize to Fit uses an “intelligent adaptive bicubic scaling which automatically adjusts for the increase or decrease in size.” So that would be similar to applying Bicubic scaling in Photoshop, but with some adaptive size adjustment too. If you also enabled Output Sharpening, then that will add another change to the final appearance based on what you chose for its viewing medium.

Scale slider in the Transform panel. I wasn’t able to find out what scaling algorithm is used here, but it’s probably safe to guess that it’s some form of bicubic.

Super Resolution. This is supposed to be based on machine learning, so its results shouldn’t look exactly the same as the other methods you mentioned. Instead, applying machine learning method should invent plausible details that the older algorithmic methods like Bicubic would not. The quality should depend on how well the machine learning model was trained.

And of course, ditch Lightroom Classic and switch to Photoshop for cropping and resizing.

Isn't the difference between these three methods obvious? Does this mean that all three solutions are equivalent in terms of processing by the software?

By @AL1M

Well, when you want to compare to Photoshop, the next question is, which of the many methods did you use in Photoshop?

Crop tool. There’s no resampling unless you enter a Resolution value. I’m not sure what resampling method is used but it’s probably some form of Bicubic.

Image > Image Size command. How the result looks depends on if you enabled Resample, and if you did, which resampling method you chose from the eight options in the Resample menu. The core options like Bicubic, Bilinear, and Nearest Neighbor go back to the earliest versions of Photoshop, so they are the oldest and simplest code. Which work OK, but the flaws have become so evident over the years that Adobe added the newer options and also pursued the latest scaling methods based on machine learning.

Edit > Transform submenu or Edit > Free Transform command. How the result looks depends on which resampling method you chose from the six options on the Interpolation menu in the options bar.

Various commands on the File > Export submenu. These let you state new dimensions for the exported copy of the image, and when you do, they probably apply some form of bicubic resampling. Some, like Save for Web (Legacy), let you choose from multiple resampling options. 

Filter > Lens Distortion. The Custom tab has a Scaling option, because this feature is an earlier version of the Transform/Geometry panel in Lightroom Classic/Camera Raw. I would guess the method is something like Bicubic.

Filter > Neural Filters / Super Zoom. This is kind of like Super Resolution in that it’s supposed to be powered by machine learning, so in theory the result should look better than the above options. To preserve the entire image make sure the Output menu is set to New Document.

Super Resolution. In Photoshop this option is in Adobe Camera Raw (but not Camera Raw Filter), and should use the same machine-learning-based scaling code as Super Resolution in Lightroom Classic. 

(Photoshop probably has a few more places to resize that I forgot about. Oh yeah, like File > Scripts > Image Processor.)

So there are lots of possible ways to do it across Lightroom Classic and Photoshop, resulting in a wide range of results. This also shows that Lightroom Classic and Photoshop have very different sets of scaling features, with some overlap in their feature sets. The many methods offered by both Lightroom Classic and Photoshop represent the evolution of resizing methods during the more than a third of a century since Photoshop 1.0 was released. The newer methods powered by machine learning/AI are supposed to look better than the older, simpler algorithmic methods from the 1990s. 

There is one more thing…the scaling methods that work with raw files can potentially look better than methods that only work with images that have been demosaiced to channels, because how you demosaic is another opportunity to control and improve how the original image detail is interpreted and upscaled, and it’s earlier in the image processing pipeline. This is part of the reason Lightroom and Camera Raw offer a Raw Details option next to the Super Resolution option.

When you put all that together, no, it is definitely not true that “all…are equivalent in terms of processing.”

In general, the older algorithmic methods can do a good job, especially the newer adaptive versions of Bicubic. Compared to that, the latest machine-learning methods can produce better-looking details at higher scaling amounts, but tend to need a good GPU and more time to process.

[AL1M]

OK Thanks.