Adobe just released a new update to Camera Raw with a feature they call Super Resolution (this should be available in Lightroom soon). The main point of this update is to use AI and machine learning techniques to learn how to upscale an image to a higher resolution. In traditional upscaling methods, one basically stretches the image to cover more pixels and it works much like curve fitting in algebra (i.e. interpolation). The difference with AI methods is that they can actually manufacture details and textures and not just scale what exists.

The obvious intent of this update is to compete with programs like Topaz GigaPixel AI for upscaling. Topaz products have become quite popular to fix less than perfect image captures and I know many photographers use Topaz at least occaisonally.

Because I like to make large prints and I have a bit of OCD when it comes to pixel peeping both these products are of great interest to me (plus they are very cool uses of technology). So I decided to run some tests on Adobe’s Super Resolution and compare it to both Gigapixel AI and bicubic smoother, which I think is the best default in the LR/PS world.

Note 1: it’s always difficult to compare upscaling methods because they often include different amounts of sharpening. Bicubic smoother may appear initially worse than other methods, but it can be sharpened more (because none was added in the interpolation process).

Note 2: I recommend you view the sample images on a laptop or desktop monitor and not your phone. For regular monitors, set the viewing resolution to 100%. For retina or high DPI monitors, view at 200%. Make sure that your browser is not scaling the images (even better is to download and view in a proper image editor). Click on the crops to view them full-screen. You can also open them in a new window by using ctrl click (pc) or command click (mac).

Test Setup

The test images were all taken from raw files in my library. Unless otherwise noted, capture sharpening was set to amount 25, radius 1, detail 25, masking 0. Some Lightroom edits have been applied but these are not finished images.

Adobe Super Resolution has no parameters. It exactly doubles the resolution by scaling the image 2x in each dimension.

Gigapixel AI (5.4.5) was set to scale the image exactly 2x. Depending on the subject matter, I used either standard or architectural mode. I used the auto button to determine the surpress noise and remove blur settings.

Close-up images are in the order of bicubic smoother (BS), Adobe super resolution (SR), and Topaz Gigapixel AI (GP). They should all be exactly 1800x600 px in dimensions. These were saved as jpeg from Lightroom using quality 65 and no output sharpening.

Animals

Sony a7r4, ISO 320, 400mm, f/5.6, 1/1600s
Gigapixel: standard, 40 surpress noise, 60 remove blur

This is an image of a California scrub jay taken in full sun. The bird is not perfectly sharp in the original raw file – I do have a sharper capture but purposely chose this version for the slight softness on the bird.

SR is crisper than GP but the white feathers beneath the eye are on the edge of showing artifacts. GP has completely removed all noise from the blue sky.

This is a close up of the feather detail. I’m not a bird photographer but to my eye both SR and GP look very good. SR has a small color shift and is a bit more red.

Canon 5D mark II, ISO 1600, 300mm, f/5.6, 1/15s
Sharpening: amount 35, radius 1, detail 100, masking 0
Gigapixel: standard, 60 surpress noise, 40 remove blur

This photo of a deer was taken at dusk when light levels were dropping rapidly.

SR generates green splotches on the fur and is unusable in my opinion. GP does not appear significantly better than BS to my eye (once you consider BS can be sharpened further).

Cityscapes & Architecture

Sony a7R2, ISO 200, 24mm, f/11, 4s, compressed raw
Gigapixel: architectural, 20 surpress noise, 80 remove blur

I processed this image using both the standard mode and architectural mode in Gigapixel AI. The architecture mode was a bit crisper so I used that for comparison.

GP exhibits some artifacts on the building details near the top of the crop (these appear in both the architectural and standard mode).

Detail appears similar between SR and GP.

Note the weird noise pattern around the door on the right in GP. This also occurs in the standard mode for GP. Gigapixel AI also overemphasized the halo around the glass panel.

Gigapixel AI produces unsual artifacts around the street light and palm leaves in the center. This also occurs in the standard mode for GP.

Landscape

Sony a7R4, ISO 100, 16mm, f/4, 1/5s, compressed raw
Gigapixel: standard, 20 surpress noise, 80 remove blur

This image is a bracket for the foreground and was intended to be part of an exposure blend. However, I mistakenly had the aperture set to only f/4 resulting in not enough DOF for a sharp foreground and background.

Both SR and GP look good in this region. GP is a bit more contrasty and has brighter highlights.

The original image is very soft in this area but GP does a better job here compared to super resolution.

Gigapixel AI result here looks very strange. There are artifacts in the water and the eelgrass on the rocks looks simultaneously sharp and out of focus.

Textures

Sony a7R2, ISO 100, 20mm, f/10, 1/20s, compressed raw
Gigapixel: standard, 20 surpress noise, 40 remove blur

This is a photo of some old rusted cars in a junk yard. I thought this would be a good test for textures.

I give the edge to SR but I also think it has a bit more sharpening applied than GP.

Prefer the GP result and the area around the hole in the window looks very good.

Both SR and GP are very good. GP has a bit more noise reduction in the green foliage which appears perhaps a bit too smooth.

Stars & Landscape Astro

Sony a7R2, ISO 1600, 25mm, f/2.8, 15s, uncompressed raw
Gigapixel: standard, 60 surpress noise, 60 remove blur

This is a single exposure of the night sky in Jacumba.

Super resolution produces some color artifacts and makes the stars more saturated and some are shifted to green. Both SR and GP appear to have more noise and the area between stars (dust clouds) look worse than bicubic smoother which would be my preference here.

Moon

Sony a7R4, ISO 100, 400mm, f/5.6, 1/125s, compressed raw
Gigapixel: standard, 40 surpress noise, 60 remove blur

This photo of the moon was taken in late afternoon.

In some areas such as the craters and blue sky, GP removes all noise making the image look a bit plasticy. However this could be mitigated by adding a controlled amount of grain or noise (this is a technique in printing that can make the image appear sharper and more detailed to the human eye). On SR the noise pattern looks a bit wormy for lack of a better word.

Sony a7R4, ISO 100, 400mm, f/5.6, 1/125s, compressed raw
Gigapixel: architectural, 40 surpress noise, 80 remove blur

This is a photo of the full moon rising and was taken just after sunset. For Gigapixel AI, I originally processed this in standard mode but noticed some artifacts in the building details which did not occur in architectural mode.

There’s not much difference in the moon itself, this may be due to it being slightly soft in the original capture. The top of the building appears crisper in GP compared to SR but there are artifacts at the tip in the support beams for the light (also occur in Gigapixel AI standard mode).

GP looks cleaner and better than SR.

In this crop GP is mostly cleaner and crisper than SR. However GP creates two different artifacts in the image. First, GP added lines to the text on the boat for “San Diego”, the result is very crisp but this is false detail. Second, the noise on the back of the boat is uneven. The left is clean and smooth whereas the right looks like dirt was stuck to the surface of the boat. SR is not as clean as GP but it also does not create these artifacts.

Summary

Both Adobe’s super resolution and Gigipixel AI seem to work fairly well. Sometimes I prefer the result of one over the other but both can have issues and generate artifacts. So regardless of which you pick, you need to check your file at 100% and mask out problem areas.

Super resolution is a bit easier to use in terms of workflow:

• it is built into Camera Raw so no additional software is needed (soon to be added to Lightroom)
• it is simpler with no parameters that need to be tweaked
• it is much faster taking time in seconds (as opposed to multiple minutes for GP).

This is too small a sample to be conclusive, but here’s how I see their relative strengths so far:

Super Resolution

• can sometimes do better than GP on fine details and textures
• has even noise distribution and there no regions left complete flat to appear plasticy
• on low ISO shots, I feel that SR is less likely than GP to generate objectionable artifacts or false details
• did poorly on the two high ISO shots and created color noise/artifacts,

Gigapixel AI

• GP does a better job when regions are soft initially
• GP is more aggressive about identifying areas with no noise and eliminating it from the scene. This is both a positive and a negative. When correctly done it results in a very good output but incorrectly done removes detail and causes uneven splotches which looks odd (I consider this an artifact).
• is more prone than SR to generate artifacts at low ISO and create completely false detail

I’ll keep adding examples to this article as I run more tests on different types of images.

As a final note, I find it hard to see differences between SR and GP when I look at the images on my retina screen at 100% which is about 220 ppi. This is a good resolution for printing and very high quality output is possible. Looking at 200% (or around 110 ppi) magnifies the difference between the methods but these will only be noticeable on the very largest of prints and if you look at it closely. With newer camera bodies that have high resolution sensors, like my a7r4, an upscaled image (2x) viewed at 110 ppi would yield a print that is an incredible 172" on the long side.