It’s headline news at the moment. AI has made massive advances in the last year and if the hype is to believed it’s coming to take all our jobs. It’s already having an impact on the creative arts and is one of the reasons Hollywood writers and actors have gone on strike.
Members of the photography community are already nervous. An AI generated image won a prestigious competition earlier this year (which was retracted when the ruse was revealed). So when I heard Adobe’s Photoshop had released an AI powered tool (currently in Beta) I had to try it on some underwater images.
The tool, Generative Fill, allows you to add, remove or extend pictures using the power of AI. You can see how to use it here.
Surface Landscape
I chose this basic shot of a Maldivian jetty to start with. I reasoned this should be the simplest for the AI to interpret and the goal here was to extend the image.
Score: 8/10
I’m impressed. Photoshop has done a good job at matching the horizon and the colour gradient through the water. It also understood the jetty texture and perspective approaching the lens. On closer inspection you can see a subtle drop in contrast and a more significant drop in sharpness in the AI generated area. On first glance I don’t think you’d tell that the AI had been used.
School of Trevally
Now to test the ability to extend an underwater image. I thought I’d start easy, so I picked a shoal of Trevally to see how the AI would deal with an irregular but repeating pattern of fish.
Score: 6/10
On initial viewing it looks good. It’s done well to imitate the natural/random distribution of the fish and it doesn’t look like it’s purely copied and pasted. It clearly understands the colour and shape of the fish it’s trying to reproduce. However, it falls down on the detail. Many of the fish lack eyes or pectoral fins. It’s not a bad try though.
Removing a Diver
Now to try and remove an object, in this case a diver, and let AI fill in the hole.
Score 3/10
This is disappointing. Of all the tasks I thought Generative Fill would perform best at removing objects. I’d assumed that filling a plain background would be straightforward but photoshop really struggled. I repeated the process a number of times to see if I could get any better results. I’m not sure why it resulted in such an obvious fringing effect around where the diver was.
Adding a Diver
Generative Fill allows you to select an area and then type in a object you want inserted. I picked a scenic underwater landscape that could be improved with a diver to draw the eye.
Score: 1/10
Lovecraftian nightmare fuel. I re-ran the generation multiple times and this was still the best image. AI is known to struggle with hands so I guess a scuba diver with fins, hoses and equipment is a step too far. The only positive is it matched the colour balance of the original image fine.
Final Thoughts
It has to be noted that with enough time and skill all these effects can be bettered manually. Unsurprisingly the tool is far better at dealing with above water images and it was genuinely impressive. Underwater, there’s still clearly a lot of development to do with the results ranging from OK to horrifying. The tool is not fully released and only available to test so there may be improvements ahead of its formal release.
New AI tools are coming and they will get better but for now at least underwater photographers can breathe a sigh of relief.
