One of the latest buzzwords in tech and editing has been AI. While the earliest implementations were a bit of a gimmick, powerful tools and developments from companies like Adobe, NVIDIA, and Luminar have brought AI to the end-user in a meaningful way. Photoshop’s March 2021 update introduces a new AI tool that promises massive resolution improvements for any camera. Does it live up to the hype?
If you’re not caught up on Photoshop’s notes, Adobe Camera Raw has received a new feature called Super Resolution. Currently available in Camera Raw 13.2 and coming soon to Lightroom and Lightroom Classic, Super Resolution uses a machine learning model to “intelligently enlarge photos while maintaining clean edges and preserving important details.” In practice, it’s a one-click way to quadruple your photo’s megapixel count, while retaining much more detail than “dumb” upscaling, like bicubic and nearest neighbor methods.
Photoshop Super Resolution Versus Topaz Gigapixel AI
This isn’t a new idea, however. AI upscaling has been around for a while. It’s even been implemented into existing consumer products, most notably in Topaz’s Gigapixel AI software. Gigapixel is built around a similar principle: train a machine learning model with a set of low- and high-resolution image pairs so that the computer eventually learns what a low-resolution area could look like in high resolution. This model can then up-sample the photos and “create” details to fill in the blanks. Depending on how this is implemented, it can be pretty computationally intense, relying on your computer’s GPU to perform a lot of work.
While there are some differences between how these programs work, with PS’s feature integrated into ACR instead of being a standalone program and Gigapixel offering some more options for customizing the processing, the end results are perfectly comparable.
For these tests, I wanted to take a look at a couple of different types of images that I often shoot and that often benefit from more resolution. To compare, I grabbed some raw files from my Mavic Air 2 (to represent aerial shots) and my Nikon Z 7 (representing architecture and product photography, as well as higher resolution). While these files are unprocessed, each program handled them slightly differently, the most important distinction being ACR automatically applying lens corrections. This resulted in a slight difference between FOV and brightness between the files, but I’m not really considering that relevant in the comparison, as you could pass a processed file through Gigapixel without a meaningful difference. Also, in the following images, the Photoshop Super Resolution version will be on the left, with Gigapixel's on the right.
Increasing Resolution of Drone Photos
In my mind, this is the worst-case scenario for upscaling. While the Mavic Air 2’s files are very impressive for a camera that can fly at 40 mph, they aren’t gorgeous at a pixel level. They can be a bit noisy even at low ISOs, and the Quad Bayer sensor, like Fuji’s X-Trans, has historically had problems with some demosaicing processes.
One thing that stuck out to me when reading about Super Resolution was how it included ACR’s Enhance Details processing step by default. Enhance Details was an earlier foray into ML-powered tools and offered a way to demosaic raw files with fewer resulting artifacts. It’s a very minor improvement in many cases, but I’ve found that it can help in cases of tricky moire, or with atypical sensor setups like X-Trans or Quad Bayer. As a result, I don’t do it by default but appreciate that it’s available.
This combination of improved image quality and increased resolution make Super Resolution seem like a very promising option for use with drones, and I can say it really delivers.
First, let’s talk about processing time and workflow. Loading the raw file into Photoshop, then right-clicking the image and selecting Enhance brings up the relevant menu. From here, a preview is quickly generated, and a new DNG can be created. Via this flow, you still have access to the same features you’d have if you’re processing the raw file normally and can also quickly see what benefit Super Resolution will offer.
With Gigapixel, loading the raw file and setting things up is quite a bit slower. There’s a delay as the preview generates, a significant delay each time you scroll or change an option as it redraws, and finally a very significant difference in actual processing times. Super Resolution produced a finished file in 3 seconds, while Gigapixel AI took 1 minute and 23 seconds.
As for the finished files, Photoshop’s version is significantly better. Two major improvements are visible. The first is an area that has been a problem for many other software tools when dealing with Quad Bayer or X-Trans files: “wormy” looking green areas. In Gigapixel’s version, there’s a very watercolor-y, unnatural look to this area of foliage.
The second major improvement is the relative absence of major artifacts in Photoshop’s version. To personify it, Gigapixel is overly aggressive in “making up” details. It creates faint patterns in areas that should be plain texture and generates noticeable artifacts in areas like text and faces. Photoshop, meanwhile, appears to just deliver a very good upscale. The drone shot, after processing, becomes a 48-megapixel shot. While it isn’t going to match a DSLR for microcontrast and sharpness, it’s surprisingly close and a drastic improvement from the original 12-megapiel shot.
The Best Option for Upscaling Architectural Images
While my Z 7 offers great resolution with its 45-megapixel sensor, more is always better. To that end, I was curious how these two scaling methods would work with a file offering a mix of organic shapes and straight lines, along with some fine details.
From this test file, I observed a similar pattern in usability, but to an even greater degree. Photoshop rendered a finished file in 6 seconds, while Gigapixel took 5 minutes and 1 second to finish its version.
Comparing the two files, Photoshop again delivered a surprisingly neutral file. There are no major problem areas, and the files still have quite a bit of “bite” at the pixel level. As Photoshop auto-applies lens corrections, the FOV is a bit different, but I think these corrections would need to be applied anyway to Gigapixel’s file, as there’s noticeable distortion present in the buildings. At the pixel level, PS’s version only has a slight issue with some fine details, like the stars on the flags. Photoshop renders them as stars, but with a bit of false color creeping in. In Gigapixel’s version, these are unrecognizable smudges as well as artifacts from false color.
Gigapixel also runs into that watercolor problem again along the chain-link fence. Here, Photoshop renders the fence as expected, while Gigapixel’s version is smeary, with individual strands of the fence almost seeming to blur out of focus.
In the architectural details, both are competent. Photoshop seems to err on the side of preserving a bit more noise and texture, while Gigapixel smooths things to a greater extent, but I think you could push either file to the same place with a bit of sharpening and noise reduction.
For about $100, I just can’t see the value in Topaz’s Gigapixel AI product for my workflow now that Adobe’s Super Resolution is available. In my testing across the range of subjects I shoot, Super Resolution delivered equal or better results in every case. Architecture, landscapes, nightscapes, product photos, aerial shots, and more all came out better in Super Resolution. That isn’t even considering the significant workflow benefits: Super Resolution is built-in to Photoshop, respects the existing ACR workflow better, and is anywhere from 20 to 50 times faster to process. If you haven’t tried out Super Resolution yet, definitely give it a try!