Adobe having been using AI, Artificial Intelligence, to speed up and improve aspects of their suite of software programmes for time now. The core of this is Adobe Sensei, Adobe’s in-house artificial intelligence technology. This has powered things such as predictive text titles in Adobe Premier Pro video editor, and content aware fill in Photoshop. Lots of technical terms that few of us fully understand are thrown around, e.g. neural networks, machine learning and deep learning. My own understanding is limited, so I’m not going to dwell on the details of the process. Suffice to say that machine learning, deep learning and neural networks are all essentially subsets of A.I., where software mimics (to a degree) the human thought process of analysing and learning. A.I. comes in to its own where impossibly large amounts of data or permutations are analysed at a speed beyond the capabilities of the smartest scientists. For us mere mortals, we can but gaze upon the results with similar awe and wonderment that neanderthals viewed comet trails and solar eclipses. So, going back to Photoshop’s Enhance, I’m going to concentrate on what we care more about. I’m going to look at the results. Adobe Sensi now comes to Photoshop’s Camera Raw processing software as ‘Enhance’, where, Adobe claims, it will produce more accurate rendering of edges and details, with fewer artifacts such as false colors and moiré patterns. Enhance also enables one to double the linear resolution of an image (i.e. 4 times the number of pixels) with, Adobe claims, superior results to previous methods. If so this can provide huge benefits. All that back catalogue of older images we took 10-15 years or so back, on 10 megapixel cameras, can be up-scaled to 40 megapixels. Images from crop sensor or micro four thirds cameras can be up-scaled to rival the latest full-frame DSLR images. Sounds too good to be true. But the bottom line is, how good are the actual results?
I first tried this out on a raw file from an older, micro four thirds camera, a 16 megapixel Panasonic GH4. I chose an image I had taken in Antarctica earlier this year. A picture of an old wooden whaling boat and whale bones, all lying in rather dirty, textured snow. The image had lots of detail and lots of texture in it, so I thought it a good test. For comparison, I processed the file in my normal way, using both Photoshop and Camera Raw adjustments. I then resampled the image in Photoshop, up-sizing to
Up-sizing an image using Photoshop’s standard image resampling Preserve Details
To enlarge using Photoshop’s standard up-sizing we select (from the top-bar menu) image>image size, then in the dialogue box that opens set the units to pixels, make sure the resample box is checked and, in the adjacent drop-down menu, select Preserve Details (enlargement). Don’t worry about the resolution; this really only matters when printing an image or adding text.
Up-sizing an image using Photoshop’s A.I. based Enhance
For up-sizing using Enhance, we need to be using the Camera Raw plugin, which automatically opens as a separate window when we tray to open RAW images. Note that the Camera Raw filter in Photoshop (which can be used for JPEG and other formats) does not give you the Enhance option. (It is possible to change the Camera Raw settings in Preferences so that TIFF and JPEG are also opened initially in Camera Raw, but this is not the default). In any event, if we shoot in RAW it is always best to use RAW here in order to utilise the maximum amount of data.
Up-sizing an image using Photoshop’s image resampling Preserve Details 2.0
But hang on! There is a third option. Photoshop have also recently introduced Preserve Details 2.0. If you are using the latest version of Photoshop (at the time of writing 24.1.1) this should be turned on automatically, but if not then you need to go into Preferences and find Technology Previews. Preserve Details 2.0 is listed there, along with a check box to enable it.
The preserve Details 2.0 then appears in the Resample drop-down menu.
Comparing the results of all three methods
With all three methods I doubled the linear resolution of the image, up-sizing a 4608 x 3456 pixel image to 9216 x 6912 pixels or, to put it another way, increased the image size from 16 megapixels to 64 megapixels (a little rounding to whole numbers). This is a major increase in image size, so what does it actually look like? The only meaningful way to compare is to look at cropped sections of the image enlarged to 100%. First I chose a crop of an area with a fair amount of detail and hard edges to see how that would work out. The area I chose was a crop of the exposed ribs of the boat, including the nails sticking out, but it also includes areas of dirty, textured snow (it’s snow with penguin footprints, penguin poo and algae in it). The snow has some texture and colour variations but lacks hard edges and I was curious to see how each process dealt with that.
Preserve Details vs Enhance vs Preserve Details 2.0
I think it’s clear there there is more detail and crisper edges in the up-sized image using Enhance when compared to the standard Preserve Details. There’s also less colour fringing around the nails protruding from the ribs (less, but not absent). The texture of the wooden ribs is markedly muddier using Photoshop’s standard resampling. Preserve Details 2.0 shows distinct improvement over the standard version and is probably about midway between standard and Enhance when considering these two aspects. However it’s not all positive for the A.I. based Enhance. The added detail in the wood also clearly contains non-real artifacts created in the up-scaling, including noticeable flecks of false colour. Some of this could be removed in further processing, but possibly at the expense of some of the detail gained. These artifacts become much more obvious when one looks at the snow in the bottom right of the image. In an attempt to add detail to the enlarged image Enhance has created non-real repeating patterns than show up as dark, sinuous ‘spaghetti’ lines that are slightly fringed with green and magenta. This creates a quite false look to the snow when viewed at this magnification. A.I. based image enhancement works by looking at very large numbers of similar images and so ‘learning’ what certain features look like. So to be fair, dirty, textured snow flecked with penguin poo does not, in all probability, offer a huge database of comparable images.
My take-away points from this are:
When it comes to sharp detail and hard-edged objects Enhance is a clear winner compared to both versions of Photoshop’s standard image resampling: Preserve Details.
When it comes to fuzzier detail and objects or features that may be uncommon, Enhance still struggles and may tend to create artificial features and patterns.
Enhance is still a very new feature. A.I. based image improvements are only going to get better as the technology matures.
Fine Art Landscape and Wildlife Prints
If you enjoyed this article why not check out my other blogs. I write about travel, the environment, marine biology, diving and wildlife. You may also want to check out my photographic prints. These can be viewed on my main site www.colinmunrophotography.com and include landscapes from around the World, people of the World, marine life and other wildlife. The prints are only available through my website, each one individual processed and made to order. This includes fine art giclee prints, limited edition prints and canvas wraps. I use only the best, carefully selected print houses employing the finest papers and printing processes to ensure image fidelity and longevity.