AI Denoise Isn't Generative AI: Debunking the Misconception

26 Oct

Screenshot of Lightroom wioth the text: Denoise AI, no it doesn't make your photo AI slop

For many photographers and creatives, “AI” has become something of a dirty word. I’ll be honest, I’m a bit of an AI sceptic myself, especially when it comes to generative AI. Concerns around “slop”, intellectual property theft during training, and environmental issues are perfectly valid concerns. However, AI is a broad term and not all technology that falls under the “artificial intelligence” umbrella is the same. This has led some people to avoid or criticise tools that have little in common with image generators. AI denoising, like the Denoise feature in Lightroom, is one of the most common examples. It carries the same label, but it works differently and does a very different job.

I’ve actually covered Denoise quite a bit on this blog and on my YouTube channel. Recently, someone left an irate comment on one of my videos complaining about how Denoise was just slop and how it wasn’t a real image afterwards. They complained that if you use Denoise, why even bother taking a photo in the first place? I was actually taken aback by this because I had assumed people understood the difference. It’s easy to dismiss such comments as someone trolling, but with such anti-AI sentiment out there these days, it occurs to me that maybe some people think that AI Denoise tools are using generative AI and are replacing your image with an entirely synthesised one.

But this isn’t what happens at all.

Traditional noise reduction mostly worked by blurring neighbouring pixels together. That removes the grain but also takes fine detail with it, leaving skin waxy and textures mushy. I mean, the algorithms are a bit more complicated than that, but that’s the gist of it. AI denoising uses a neural network trained on large numbers of image pairs, each showing the same scene with and without noise. When you run it on your photo, it looks at each area and estimates what the clean version of those exact pixels most likely was. It removes the speckle and keeps the edges and textures that were really there. Lightroom's version even works on the raw file before it becomes a normal image, which is part of why the results look so natural.

In other words, it uses machine learning and statistical analysis based on thousands of images to figure out how to adjust your image on a pixel-by-pixel basis to achieve the best noise reduction. This is a very simplified version of what is happening. Adobe’s own blog has more details on how their Denoise works, but most other tools that have similar technology will work in the same way.

But the important thing to note here is that it is NOT using generative AI.

So what’s the difference?

Generative AI creates content that was never captured. Ask a tool like Generative Fill to add a sky or remove a person, and it invents new pixels based on what it has learned images usually look like. A denoise tool’s job is to improve your image: it modifies your image rather than creating a new one. A generative tool's job is creation: it is free to make up something plausible where there was nothing. You’ll often see this in more recent tools like Lightroom’s (cloud) prompt-to-edit feature, where it tries to edit your image, but because it uses generative AI, it’s effectively creating a new version. This often leads to different details in your image, so people might appear where none were there before, and some other fine details might completely change. It’s perfectly ok, in my opinion, to be mad at this kind of thing. But that’s not how AI Denoise works at all.

There is some crossover, of course, and they are built on similar technology such as machine learning and neural networks, but that’s where the comparison really ends. Denoise isn’t perfect, and can occasionally result in artefacts and unnatural-looking images. Still, a denoised photo is your photo, just cleaner. Or to put it another way, denoising edits your existing image data, whereas generative AI creates entirely new image data. Even if you’re using it to edit your photo, it’s still a new image.

AI Denoise isn’t new either. It’s been around for a while now. Adobe introduced it in 2023, and many other software companies now have also added similar features. In particular, there is DXO’s Pure RAW, which uses DXO’s own flavour of denoising technology in combination with its popular lens correction profiles, and Capture One also recently added AI denoising to its toolset. Other notable mentions are Luminar and ON1 No Noise AI.

I hope, like Denoise itself, this has made things a bit clearer. (Nerdy and terrible Denoise joke!)