Over the past year or so I’ve continually been subject to a barrage of Facebook ads for some photo enhancement software called ‘Photo AI’ from Topaz Labs, almost all of which contained before-and-after pictures with seemingly incredible results. The idea being that you could feed in fuzzy, low-quality images, and somehow sharpen, upscale, and denoise them into masterpieces. As an experienced photographer and techy-sort, this all struck me as completely unbelievable. You simply can’t scale up an image reliably in the way that they were claiming. Right?
Curiousity eventually got the better of me, and I downloaded the trial version – prepared to scoff and shake my head at the results. However, that isn’t actually what ended up happening. Instead, I bought a copy. So eh, am I just a sucker? Possibly, but let’s take a look.
Psst you will want to click through to the full size images to get the best view on the differences.
Here’s what the UI of the Mac desktop app looks like:
You have your original image on the left, and the enhanced image on the right. There is a sidebar with a few different options, including upscale, de-noise, sharpen, face recovery, and text preservation. Each of these expands out to give you finer control over the individual parameters, though it isn’t immediately clear what exactly they do, or how they interact together.
When you load in an image at first, Photo AI analyses it and applies certain settings on auto-pilot which you can then change to suit your tastes.
Now, I’ve been running this blog for over a decade at this point, and in that time, camera technology has dramatically improved. That means that a lot of the pictures I shared on here back in 2010 now look pretty crap – their imperfections highlighted by the increased resolution of modern screens. Part of the reason I was interested in trying out Photo AI was to see if I could upscale, sharpen, and generally rescue some of the older images that I have a lot of fondness for, but which don’t quite cut the mustard in 2023 – effectively breathing new life into them. This may well seem like a fool’s errand, but let’s take a look at some examples.
The image below is a portrait of my friend Jenn, that I shot years ago on an old 25mm f0.95 lens, which is notoriously difficult to focus. Despite liking the picture itself, the result is a bit soft, and I’ve often found myself wishing it had been a bit sharper.
I decided to use this as a test for the Photo AI software, chucking in the web resolution version, which came in at 1400x935px, to see how it handled it. Here’s the result:
Now this isn’t perfect by any means, but the difference is pretty incredible – not least because the new picture is now 4242×2833 pixels. That’s more than three times the size of the original. For a first go with an app that I had written off as snake oil, I couldn’t quite believe how well it had turned out.
I had to try some more.
Here are a couple other tests, which I had to screenshot from within the app itself, as I hadn’t yet bought a license.
This picture was taken in Alabama back in 2011 or thereabouts – one of me and a long lost pal Melissa. I’ve always really loved this picture, but the original (on the left) is really noisy, in a way which i think detracts a bit from the image itself. After a few rounds of tweaks with Photo AI, this is how the comparison stacked up. It’s not perfect, with the noise distribution feeling a bit un-natural, but it showed enough promise to pique my interest.
This time, a shot of my dog BMO – that I took a few weeks back with my GF1 and that 25mm f0.95 lens again. The image on the right has been enhanced with AI, and has managed to reduce the noise and sharpen his face just enough to make it stand out a bit more, which I think looks pretty good.
With any AI tool, there comes a point where even though the results might seem really impressive, the outcome is somehow hyper-real, to an uncomfortable degree. Here’s another picture of Melissa which I cleaned up using Photo AI:
I like that there’s a lot more clarity in her face here, and a focus on the eyes since the noise has been reduced, but it also doesn’t really quite look like her any more. It’s too perfect. It produces the same unsettled feeling in me that you get when you come across a computer game character which looks human, but isn’t quite… and you don’t know why. The AI is doing far too much heavy lifting; taking things too far.
Here’s another example.
If we take a look at the below image of my pal Chris that I shot with my GF1 while on tour with his band in Aberdeen back in 2011 or so, there’s a fair bit of digital noise. It’s not awful, but it’s also the kind of thing that I wouldn’t expect to see from modern cameras. Let’s chuck it into Photo AI and see how it gets on…
As you can see, the noise in the picture has almost completely gone, which is pretty amazing, but everything is too smooth now. There’s no grit or character. No texture to his jacket. It feels more like a painting, or something that’s been generated in MidJourney.
This was just on the default, suggested settings though. After fiddling about a bit more, I got this:
This one is definitely better, though something still doesn’t seem quite right. It could be that the pattern of noise doesn’t match what I would expect as a photographer, or maybe it’s just the case that I’ve been staring at the original for far too long, and nobody else would notice – but there’s still something amiss. If I had to pick out of the three of these, I would still go with the first.
Let’s try another example.
Again, while the result here from Photo AI’s defaults isn’t terrible, there’s something a bit off about it. The sharpness of the eyes versus the smoothness of the picture doesn’t add up, even with a super-fast lens like the 25mm f0.95. I messed about with the settings for a bit, but didn’t come up with anything I liked. The results were too perfect. Eerily so in fact.
Noise isn’t always bad
One of the main reasons I was interested in Photo AI was to reduce the harsh digital noise that was present in a lot of my older pictures, before higher ISO technology improved. However, as I experimented I realised that the noise often added more than it took away.
Here’s an example. This time, I’m really going to stress it out, with an extremely noisey portrait that I shot of a guy called Nick at a wedding years back – again with my GF1.
Jeepers. It really has obliterated all of that noise. Again though, something about the smoothness of the texture on his jacket just doesn’t fit. As somebody who is finely attuned to lens blur and depth of field… I can tell immediately that we’re not looking at that. I suspect part of the issue is that the original image is already pretty bold – deliberately edited that way. The noise gives the shot texture that balances out the contrast. Without it there, it doesn’t seem real. Let’s try dial in the controls a bit more finely.
Hey, we might actually be getting somewhere. AI has sharpened up the face, allowing it to stand out, but we’ve also retained enough of the original image’s character for it to not trigger that uncanny valley feeling. It’s not quite there yet, but I think with a bit more tweaking I could get to a place where I preferred the enhanced version over the original.
Sometimes though, AI just can’t figure out what to do with particular elements in a photograph.
Here’s a shot of my wife Grace, which I took on the night we got engaged, back in 2013. Again, this was shot with the GF1 and 25mm f0.95 combo. I love this picture, but thought that maybe the sky could be a bit less noisy, and her features a bit more defined.
and here is the first run through Photo AI…
I like that Grace’s face and eyes are sharper and stand out more here, but the sky looks awful – a smudgey blur of browny yellow. No matter how much I fiddled about with the settings in Photo AI, I couldn’t get it to look any better, and I suspect that’s because… there’s nothing actually there. The sky is empty. It’s just a gradient. Even when I tried to remove the de-noise function completely, I still had issues, as the upscale engine struggled to deal with that sky.
After a lot of tweaking, I got to a place that I think looks pretty good.
Now apparently the results are meant to be better if you apply Photo AI directly to RAW files. So far, I haven’t been doing that, partly because I use Lightroom to edit my pictures – and if you want to insert Photo AI at that stage, you have to load it as a plugin, which runs separately. To make matters worse, plugins are only supported by Lightroom Classic – not CC, and so it all becomes a bit of a faff.
From my brief tests, the results seemed pretty good – but I’m not sure they are worth changing up my entire workflow for. What seems like a smarter option is to identify particularly troublesome images, and import the RAW files into the standalone Photo AI app first, then process them in Lightroom after – but you might not know which are going to be a problem until you’ve done the processing, in which case it’s too late.
All of that said, I think I’ll stick with JPEGs. All my friends are after all… Okay, okay. I know.
Overall, I have been pretty impressed with Photo AI, though I guess I shouldn’t be surprised. This kind of technology has been advancing rapidly over the past few years, and it is only going to get better. It’s actually pretty crazy to think about how good it could get.
In many ways, Photo AI is most impressive when it is deployed sparingly, as opposed to the more extreme cases. Often I found that the subtle adjustments are the most effective, and whilst you can of course over-do things easily, ending up with hideously alien-looking images, that is the case for all tools of this nature… and ultimately, that is how I view Photo AI: another utility in the toolbox to address a specific need. Having tried similar solutions from Adobe and others, this is currently light years ahead.
Below is an example of what I mean. The shot itself was fairly unremarkable, but the lady’s face lacked a bit of clarity thanks to digital noise – something that I had struggled to fix in Lightroom. With some careful adjustments in Photo AI, the output wasn’t mind blowing, but it cleared up just enough to mean that I was happier with it – and that’s what matters.
Here’s another example, where the guy on the left was just a bit too noisey for my liking.
and the cleaned up version.
You can see better in this side by side comparison. The impact is subtle, but effective.
I was sure that Photo AI would be rubbish, and I am glad to have that assumption misproven. You can’t expect to just throw any terrible picture in there and have it magically cleaned up, but does actually work, and does a pretty good job. The upscaling functionality in particular is brilliant – though it does take a bit of time to tweak the various parameters to get things looking right. I’ve only had a few hours with it, so I am sure I will get quicker and better results with experience. If anybody can tell me how to save individual images, as opposed to all at once, that’d be great. Kthx.
In an age where modern cameras have incredible low light capabilities, you might wonder what the use case is for something like Photo AI – but I can think of a bunch of times that this would come in incredibly handy – particularly for professional photographers. Even with fancy sensors, sometimes you don’t quite get the exposure right on an otherwise great shot… or maybe you just want to restore a bunch of old pictures.
The usual price of Photo AI is 199USD. I paid 159USD, but even that feels a bit steep – particularly given some of the workflow niggles, and the pace by which this technology is advancing. The license is a one-off cost, as opposed to a subscription, and includes one year’s worth of upgrades, but it seems inevitable that you would have to re-purchase in the future at some point, if only to retain compatibility with OSX versions.
If you are regularly shooting in tricky lighting situations, like to use older digital cameras, or have a lot of pictures in your archive that you want to spruce up, then Photo AI might be worth a look.
Note: There is a Topaz Labs affiliate program. I am not part of that, and paid for my own copy. They didn’t sponsor this in any way – though I am 100% open to bribery.