NZFF

Last year I posted about my experiment with AI imaging, images from text prompts, via Midjourney. I have been meaning to try this again now that it has had time to improve, but I think it cost $30 per month back then, so I only did one month.
I will check it out again one day, but in the meantime there's the new Photoshop beta tool, where you just draw a mask on your image, type a text prompt of what you want, and in 15 seconds it comes back with 3 versions to choose from.
This particular AI tool has caught Youtube by surprise, everyone is in awe of it, and I can really see why. It does require a Photoshop subscription, which I have anyway, and at the moment there are unlimited prompts during the beta, once it is out of beta it will start to cost, but Adobe have said that there'll be a number of free prompts included in the subscription. So probably cheaper than Midjourney.
Here's some experiments so far.
First, I've started with a photo I took travelling between Timaru/Dunedin when I updated Taieri. I think this was the town of Herbert, I always wanted to stop there to look around because of their cute shop, which I have a nice photo of now, I just won't show it here. When I looked around the town I noticed a lot of what looked like junk yards, and I really wanted to explore them, but there were people at these properties, and I didn't want to appear nosy... I've always like the idea of a town devoted to junk, so with this new tool I've been experimenting with a couple.
First, I had a photo of the cab of an old truck -- one one of the right of this image -- so I wanted to add some extra junk, in particular some old rusty fair rides. So I drew a mask beside the truck, and typed 'rusty carousel' as the prompt. I repeated the mask/prompt and ended up with this:


It appears to be a bug with the tool that sometimes it will ignore the prompt and give a result based on what it considers the subject, in this case the old truck. Weird.

After a few more tries, I settled on 'rusty round-a-bout' or similar, and a few different goes gave me this:


So it got it right this time around.

Here's another photo I took. Notice this is portrait format....


One thing which the tool does particularly well is extended an image, so the background to the right of the image is all AI generated based on the original image. This time the prompt was 'rusty machinery' and I have this shot now, no real point although I do like the theme!


Another shot taken when my nephews stayed with us at Diamond Harbour almost a decade ago. It originally showed the boys mock-fighting, but I had the AI tool remove them for privacy.

This is a bunker at the old barracks on the Lyttelton side of the harbour, probably used for ammo storage. I told them a story about a cult which lived there when I was a kid, and their unexplained demise, just to creep them out. Now I wanted to give an idea of what it looked like before this ficticious cult were all dismembered by persons/demon unknown...

What really astounds me is the lighting on the statues, it perfectly matches the light from the window.

I do like this better than Midjourney, because you start with your own photo. If you have a photoshop sub, grab the beta and give it a go.

WOOOOAAAHHHhhhhhhhh... that's CRAZY cool !!!!

The last one is very impressive, as you say the lighting excellent, but the theme and contents...

That is an amazing tool - but who owns the copyright on the final image? Does PS credit the owners of the images of the rusty machinery? Looks like an instant beanfeast for the legal profession.

So you thought all the pictures you uploaded "for free" were just for your entertainment?
Nope, they have been used to train up the generative AI.
Rusty machinery? Easy peasy.

It's with us, in every respect.
I do a bit of technical reporting, and often getting started is the hard part.
I idly got ChatGPT to write me bullet points on a particular topic, and after some reflection, I built the report around the result.

Like all powerful tools, there is a good chance the wrong kind of people get their hands on it, and the consequences might be quite significant.
I am hoping there is enough wisdom and expertise available to governments, to legislate appropriate curbs in timely fashion.
The people who should know, claim this is right up there with the discovery of Fire.

Thanks for the interesting post.

So, 'government wisdom' and 'government expertise' join 'military intelligence' and 'journalistic ethics' in the growing list of examples of oxymoron.

I know, let's get Elon Musk to do it!

emfrat wrote:That is an amazing tool - but who owns the copyright on the final image? Does PS credit the owners of the images of the rusty machinery? Looks like an instant beanfeast for the legal profession.

The generated image did not exist before I entered the prompt. Adobe's AI tool is 'trained' using the Adobe stock image library, which is millions of photos, all 'belonging' to Adobe.

You aren't looking at an actual machine, it's just that the AI has seen enough photos of machines to know how to approximate one. Same with the truck cab images -- neither of the two generated cabs ever existed, in this case I suspect the AI didn't even consider them to be trucks, it just looked at the shape/size/colour/tones/textures of the real cab and built a 'something'. This is something which happens often with some prompts, before I settled on the 'rusty machinery' prompt I used 'rusty carousel', and this is one of the results. It is obviously rusty, vaguely carousel-like, but the yellow/blue bit is very much based on the postbox. This object not only never existed, it really has no reason to exist!



It is scary that this year's election is our first which uses AI imagery, we've already seen it with midjourney prompts being used to create online political ads. The election after that will probably use AI-generated video, which will be another level of trouble.

Thanks Rob. Scary indeed.
Mike