AI in photography: where should it stop?

AI is a bit of a hot topic in photography right now as it re-ignites old debates about deception, fabrication, truth and lies – and whether photography can even be trusted any more. Actually, photographs have been fabricated long before this. AI has also brought new image-editing tools that are a force for good, so it’s not a simple argument. So let’s take a look at what AI in photography actually involves, and how it’s a broad spread of innovations not just one.

The first thing to do is establish a clear distinction between ‘generative AI’ and AI enhancement tools. Generative AI creates or adds things that didn’t exist in reality, while AI enhancement tools enhance what was there. In between are some grey-ish areas.

Generative AI: pure evil or revolutionary money-making opportunity?

Well, it’s probably both, depending on which side of the fence you’re on, and whether you stand to gain from this tech or stand against everything it represents.

Right now, generative AI essentially means using text and/or image prompts to create a photograph of an imaginary subject, scene or scenario. The results have become terrifyingly realistic. Different generative AI models are available including Adobe Firefly, ChatGPT and Google Gemini amongst others.

AI-generated imagery is very much in the news right now, especially where it’s used for propaganda and political gain. Tools exist to help identify AI imagery, though this doesn’t stop it spreading. We can comfort ourselves with the fact, however, that this kind of false imagery has been around for years, whether it’s been doctored in Photoshop or a video editor, or it’s footage from a different conflict or event being presented as something new. This has been going on for ever. All that’s happened is that the tools have changed.

What about ‘generative remove’ tools? These are like a subset of generative AI where, if there’s something you don’t want in an image you can have the AI remove it and substitute what you do want from, typically, a text description. Here, you’re only fabricating part of an image, not all of it.

Generative AI does have a positive role too.

Artists can use generative AI to produce images which would be difficult to create using conventional media, or which would not carry the same weight, or intention, or expressiveness. That’s an artistic decision. The key here is that art is not presented as fact. This is where AI becomes controversial – when it appears to illustrate something factual. That’s not the role of arts.

So it’s very important to distinguish fact from fiction, and also to examine the motives of any AI creator. They can be good, they can be bad.

Generative AI has also proved valuable for filmmakers and commercial content creators developing pitches or storyboards for clients. The results can be very compelling and much cheaper to pitch. You just have to hope the client doesn’t like your pitch so much they go with that and don’t hire you for the project.

Adobe users will have spotted that Adobe is now offering third-party AI models alongside its own Firefly model. That’s interesting in itself, but there is an important point here. Adobe has always advertised its generative images as ‘commercially safe’. In other words, the Firefly model will not create any image with identifiable trademarks or products that might leave you open to IPR infringements and legal action. The third-party models now being made available within Adobe apps come with no such guarantees.

Generative AI is not free: credits where credits are due

What’s that marketing term? Bait and switch? So in the early days when generative AI was experimental, we got to try it out for free. We might even have imagined it would stay free.

Not so. Now that generative AI is a mature product, AI companies have rapidly adopted a standard charging scheme based around ‘credits’. Each generative operation costs a fixed number of credits, depending on its complexity. You may get a certain number of credits built into your software subscription plan, you may need to take out an additional subscription plan to get enough credits for what you want to do. Generative AI images are not that expensive; generative AI video is fantastically expensive. As the AI models improve, expect the cost in credits to increase. They’ve got to pay for those AI datacenters somehow.

AI enhancement tools: not the same thing at all

Now a lot of people lump all image-related AI into a single thing. That’s not quite right. There are many AI tools designed to improve and enhance existing images without adding or replacing anything. These are typically on-device tools which don’t carry any additional cost in credits and don’t mean uploading your images to a datacenter somewhere. Two key examples are AI denoising and AI masking (and perhaps AI upscaling).

AI denoising uses deep learning techniques to remove or reduce noise in high-ISO RAW files as part of the RAW processing phase itself. A prime example (sorry) is DxO DeepPRIME. The AI is not introducing anything that wasn’t there in the first place, or couldn’t be inferred from the existing detail. So the AI might see there’s some fuzzy grass or animal fur in the scene and enhance/reconstruct it based on what it knows these surfaces are like. In any event, AI noise reduction is simply aiming to improve on noise reduction techniques we’ve been using for ages.

AI masking is another example of AI being used to simply and perfect a job we were doing already, by identifying skies, people, main subjects and other objects automatically and creating precise and accurate masks so that we can make better local adjustments. Again, the AI is not creating anything that wasn’t there before, simply making it easier for us to use the same editing techniques we’ve been using all along.

What about AI sky replacement? This is not generative AI. It occupies the middle ground, making it easier for use to fabricate scenes we would probably have been fabricating already in programs like Photoshop. AI sky replacement is really just AI masking with added extras.

How is AI relevant to photography and cameras?

At the moment there is little AI crossover in camera tech. Canon has experimented with in-camera noise reduction and upscaling but in reality you’re going to have much more success doing this on a computer. So far, the AI in cameras has been limited to enhancement tools. There are no cameras right now which offers generative AI in lieu of auto photographs, but no doubt that will come. There is talk that the ‘AI-native’ Caira camera will come with some kind of generative AI capability at some point.

But will generative AI affect photographers? Certainly. It leaves us with plenty to think about and it’s almost like a fork in the path where we must choose between fact and fiction and figure out exactly what it is we’re trying to present.

For professional/commercial photographers there’s even more to think about. Product photographers may find it easier to use generative AI prompts rather than actual product shots (or, worse, manufacturers might do it themselves), while stock photography is surely doomed, or at least facing a major decline. After all, if you want a custom-made royalty-free stock image to illustrate a concept, why go looking for one when you can just make one?