Monday, 10 July 2023

Infrared generative AI

I've been 'playing around' with a couple of generative artificial intelligence engines that carry out text-to-image. I was interested in how they understood the look of an infrared photograph.

This was partly because I find that at least one of my IR photos has found its way into the models used for generative AI. Without permission of course!

I tried a fairly simple shot; that of a tree. A good test for the basic Wood Effect. The results are quite good and, given how unreal IR photos can look anyway, it would be difficult to tell from the real thing. Sometimes the background detail gives things away but I found I could quickly get good results.

The two I tried are Stable Diffusion and Adobe Firefly (chosen simply because I have access to them) and the text prompt is the same for each: Infrared image of an oak tree in a meadow.

This is the Firefly result ...

As you can see, I got a false colour image without specifying, and in this case it's not a bad result. Often the system would put arbitrary colour washes over the image, but here the result is as you'd expect.

This is the Stable Diffusion result ...

Here, it's black & white straight away. Again, a pretty good result.

I did also try to generate some thermal images but neither engine could get anywhere near what you'd expect, presumably because there are not many, if any, thermal images in the training set.