The programming team behind Pixelmator Pro has just released an interesting update that introduces the ability to enhance the resolution of an image with some quite impressive results. Billed as being similar to what you often see in cheesy police dramas, the software invents detail using some sophisticated techniques.
The technology was achieved by downscaling a batch of images and then “teaching” the Pixelmator software how to upscale them back to their original resolution. As explained in this blog post, Pixelmator Pro is “predicting the values of each new pixel,” guessing the “edges, sharpness, contours, and patterns that traditional algorithms simply cannot.”
Pixelmator compares this upscaling technique to algorithms that are traditionally used to upscale images, such as Nearest Neighbour, Bilinear, and Lanczos, demonstrating the difference in quality that its ML Super Resolution offers. I decided to have a quick play myself, and the results were quite impressive.
In this test, I took a 1,920-pixel wide image and downscaled it to 800 pixels. I then resized back to 1,920 pixels using Nearest Neighbour when upscaling in Photoshop. I opened the 800-pixel file in Pixelmator Pro and repeated the process, this time upscaling it using ML Super Resolution. Click for bigger.
The original file at 1920 pixels in width.
A crop of the file after it had been resized to 800 pixels and then upscaled back to 1920 pixels using Photoshop and Nearest Neighbour.
A crop of the resulting file after upscaling it to 1920 pixels using Pixelmator Pro's ML Super Resolution.
Pixelmator is keen to point out that the process is completely secure: "We've fit the entire trained model inside the Pixelmator Pro app package so all the processing is done locally and no data ever leaves the device. This is the first such tool in a consumer image editor." Pixelmator is also emphasizing its speed and ability to remove noise and JPEG artifacts.
It certainly seems effective, and you can check out more examples in the Pixelmator blog post. What are your thoughts? Leave a comment below.
At least based off PIXImperfect's comparison, Gigapixel is bad. There's weird horizontal lines. See below vid which should start at 17:11:
https://youtu.be/5LrPKoNvPPc?t=1031
Gigapixel A.I. is _not_ bad. That, was Unmesh experienced in his video, is _not_ normal. As he says, there has to be some problem with his system (Win 10 is already a problem), the software version of Gigapixel A.I.or (most likely) the image itself.
I personally use the software quite a bit for some time and never had such phenomena like horizontal lines. I experienced mainly very good results, but also some images, where the upscaling is not possible because of unacceptable results. But in general, Gigapixel A.I. is a great tool most of the time.
I don't ever use this type of software, but the Topaz Denoise AI is truly amazing and by far the best purchase I've ever made with software. I love the results and now comfortably shoot at ISO 12,800 on my Nikon D4s bodies and get terrific results after applying the DeNoise AI software. I'm sure Topaz Labs makes great software all the way around, but I have yet to try any of their other software yet.
Dave that is brilliant... give it straight back at them mate.
Gold!
I did the exact same thing when I first read the article, and got the same results as you Dave.
The reason it's 1500px is because, from what I can tell, the site limits the upload sizes. 1500px for horizontals and 1200px for verticals.
The PS Auto is cleaner. Though, Pixelmator has more detail on the structure, but, at the expense of more artifacts all around.
I love this new ML super resolution. Used it on a full size image from an a7r3 and it churned out a 387mp image. While I’m not doing art prints currently, it’s a look into what will be available for future proofing work. Also the ability to completely remove noise without a loss in detail.
Machine learning
The Nearest Neighbor is the worst algorithm to enlarge a photographic image. It does have its place when it's used to resize pixel artwork. However, Bicubic or Lanczos should've been used to enlarge photographic images.