Updated: July 17, 2026
I go about life with a firm belief that modern software is mostly crap. This makes the joy of discovering un-crap programs that much greater. Sure, being an eternal pessimist means all your surprises ought to be good - or that you lack in negative imagination - but the thing is, I'm an optimist, believe it or not. Hence, me getting surprised and disappointed so often, despite having enough experience to know otherwise. Because I strive for quality and efficiency.
Upscayl is a rare example of software-well-done from the past decade or so. I discovered the program when I wanted to fix some old images. What I liked straight away is the numerous installation formats and options, even for Linux. I grabbed the snap, set it up on an old laptop with hybrid graphics, and looked no further. The tool did what it promised and then some, with excellent Nvidia support even on elderly hardware, plus neat, smooth cooperation with the PRIME setup, and solid results to boot. Now, let's have another look at this gem, shall we.
Small visual changes, new models #
The program offered a first-run guide on next launch. The sidebar has been rearranged in that you now get to see more detailed explanations about what the different models do. You also get images for comparison, which should help you decide what to choose. But overall, it's still the same familiar utility. The best part is that there are subtle changes here and there. All useful, mind, like the zoom-in on finished before-after results.
I let it run on the same set of images I used the last time. The system correctly decided which card to use, and how. Upscayl used 1.2-1.3 GB of VRAM. The 580.XX family of drivers, kernel 6.17, 4K display resolution, on a machine now in
its twelfth year of existence. Not bad, not bad.
Better results? #
Yes. Upscayl now also comes with the "Lite" model, which lets you get a quick preview of your work, before you decide to commit to longer renders. Furthermore, the total processing time has improved a little. Without any proper benchmarks, I would say 10-20% quicker. However, this could also be due to better graphics drivers, or maybe the optimization in the kernel, which definitely boosts the performance and responsiveness on this laptop.
I paid close attention to the fine detail in my old scanned photos. Previously, in the chalet example, as the image was quite blurry, the post-processing results were actually worse in some regards, despite the upscaling. Now, with the High Fidelity model, the level of noise seems the same, maybe a tad higher. This is quite encouraging.
If you don't zoom in too much, and just appreciate the images for what they are, you won't really be able to tell the difference. For example, the original 2764x1933px versus the enhanced 11056x7732px upscayled version, as shown in GwenView in Kubuntu 24.04. Also, the entire procedure took about 4-5 minutes. On a machine that's 12 years old, yup. Next, I looked at my kitten example. Here, the results were even more spectacular. Again, the High Fidelity model even "unblurred" some of the details on the fur. Sure, this is entropic magic, but it's actually quite decent. Considering the fact the original is only 720x540 pixels, blurry and such, being able to upscale it by 4x without any loss of quality - on the contrary, you even gain some - is quite commendable.
Limitations still apply ... #
Now, I got you enthused, so let me bring it down a notch. Upscayl does a fantastic job, but it can't do miracles. Testing all sorts of models and photos (mostly of people), I did notice certain "problems" so to speak. Again, if the signal-to-noise ratio is high, Upscayl will struggle.
Indeed, let's look at some fresh new examples. From my Polo GTI review, 4x Ultrasharp upscaling. The entire process took about 80 seconds, on battery power. This model is a bit overzealous, and does not handle sharp transitions among textures that well. It handled car details superbly, as you can see, but it messed up the grass.
Here, for example, the Standard AI model does a better job. Again, for this 800x600px image, the total processing time was roughly one minute. This seems to be a good rule of thumb for most images and models, in that you should take a 800x600 grid to be your "astronomical" unit (one minute for me), and then multiply this with the square of whatever scale factor you choose. Some models do work differently, though. More soon.
Remacri seems to be a great model. It does a good job in these mixed conditions - nature plus smooth surfaces and textures, like cars. You also get barely any change in the color spectrum, although the post-render images are slightly brighter. Likewise, the rendering time for an 800px photo took about the same.
From my
[Abarth review](../life/abarth-595-review.html) ... looking good, looking good.
For buildings and/or urban settings, like in my teaser image above, Ultramix is the model you want. But it's also much much slower, roughly 5-6 times so. When I rendered a full-sized photo of about 2500px per side, the procedure took a solid 40 minutes to complete. The much smaller 600px-to-the-side crop still took about 4-5 minutes to finish. But the results are jolly.
Lastly, Remacri showed its value for all-nature photos, too. I took one of my own, purposefully downscaled it so it looks potato-ish on the big screen, and then let Remacri do its job. This photo is a bit tricky, because it's mostly snow, high glare, but still, with some important details. And yet, the model did wonderfully, fixing and smoothing out problems with good accuracy. No weird shapes or textures. I expected the results to be worse, but nope. Since I have the full-scale original, I could actually compare the two. Very decent results.
And what about people? #
This question may intrigue you if you have lots of old photos, including degraded family albums and such. Well, from my testing, it would seem, for people, especially mid-distance photos of average quality, you usually get smeared features and details. Low-light conditions exacerbate the problem further.
If you want to upscale photos of people, the original ought to be of decent quality and size, and rather good light. Otherwise, the subjects will look like partially melted wax dolls when you zoom in. With nature, as you've seen above, the program is more forgiving, probably because no one truly focuses on each grass stalk or tree branch or such like. The same applies to buildings. The alpine chalet example represents this quite well. Look at it this way, too: when we look at humans, we expect certain things. Hence, Upscayl needs to be perfect each time, or things will look wrong. When we look at the nature, we average lots of the stuff we see, from gravel to insects to flowers to leafs to waves or similar.
Summary #
Now, here's a wee summary of my findings:
- I found High Fidelity to be either excellent or rather meh with photos of people.
- Remacri works great for mixed scenarios, and Standard is also pretty good overall.
- Ultramix is the go-to model for urban settings, but it is quite slow to render.
- Most models (except High Fidelity) usually smooth out the details, so you lose some information.
- Light (or contrast) seems to be the biggest enemy in the predicted quality of the output.
- By and large, all of the models are better than the last time.
And now, it's time to end ...
Conclusion #
Yes, I am very happy. Upscayl is a superb tool, an exception to the rule, a shining star. For those who think I'm opposed to "AI", nope, I'm not. I'm merely opposed to nonsense ideas and bad (commercial) implementations of pointless features. But when done right, the technology can do wonders. Upscayl is a prime example of quality, usefulness and privacy. The tool works fully offline, even if you disconnect from the Internet. It is also quite fast, and friendly to old hardware. The speed is decent, nonetheless, and all the models work better since my last review.
All in all, if you need some sort of tool to fix your photos, and you would like to do it locally, without going to the cloud, Upscayl delivers. The program is available for all the different software platforms and operating systems, and there's a whole range of installation formats. And if I'm not mistaken, you can add your own models, too. Did I not mention the batch mode, too, so you don't need to be sitting in front of your computer, waiting? Well, then, what's not to like. On this cheerful thought, let us say bye bye.
Cheers.