DSF

Hi all. It's a bit quiet in here at the moment so I thought I'd give you something to ponder over.

I've done a few searches and can't really see anything relating to the use of AI in producing eye-catching designs.

One program which I've played around with recently is Discord's 'Midjourney.' - https://www.midjourney.com

Basically you enter a text 'prompt' and the AI trawls billions of internet images and tries to match your prompt, with many variations. It's as simple as that! For instance ...

[ATTACH=CONFIG]6978[/ATTACH]

Above image shows someone's text prompt in Midjourney and a resulting image.


[ATTACH=CONFIG]6977[/ATTACH]

This is the actual image produced by the text prompt. Not bad for a phone case maybe.

Or maybe you need some unique wall art that is, er, off the wall! ...

[ATTACH=CONFIG]6980[/ATTACH]

[ATTACH=CONFIG]6979[/ATTACH]

See full-size image herein incredible detail!

Midjourney is fascinating, but DALL-E 2, which will soon be released, is the bees knees!

We just sat and watched the videos, fascinating!

My Son is studying game design so found this very interesting. It's sparked my imagination from a design point of view as well.

It's amazing stuff Justin. I know people who cringe at the words 'Artificial Intelligence' (possibly contemplating the 'robots are going to take over the world' scenario), but, like it or not, it's here to stay, and when it's used for the above purposes it is both practical and enlightening.

Best of all, open source software is now rapidly catching up with the main competitors and all of this will soon be available FREE of charge. We are just at the very beginning of a massive step forward and it's very exciting to be a part of it.

Lol, yes it didn't take long to see the conspiracy theorists

It is going to be a huge leap forward, better to embrace.

Now I just have to find out how to have a go!

OK, I'm on and I am immediatly convinced. My very first requests have thrown up some incredible stuff.

webtrekker;151077 wrote:
AI trawls billions of internet images

I think that is the problem!

The "program" simply trawls an internet search site e.g. Google Images and uses their algorithm to pick an image created by someone else, selects subject in the same way Photoshop and many others do and then superimposes it on your background.

It isn't actually an original as it uses (probably) copyrighted content and also probably a low resolution which looks OK on the screen but not in print.

I'll fetch my tin hat.

StellaD;151099 wrote:I think that is the problem!

The "program" simply trawls an internet search site e.g. Google Images and uses their algorithm to pick an image created by someone else, selects subject in the same way Photoshop and many others do and then superimposes it on your background.

It isn't actually an original as it uses (probably) copyrighted content and also probably a low resolution which looks OK on the screen but not in print.

I'll fetch my tin hat.

Hi Stella. I was being a bit vague when I made that statement as I didn't really want to make a long-winded post of the actual methods used to create AI images.

However, to save me the bother (ok, I'm a bloke, Im lazy!), someone else has made a long-winded Reddit post of their own which is quite a good explanation, so I'm posting it here ...

That’s not really how machine learning works. It’s not using pieces of existing images in the training data to mix them together in a “new” image. It’s learning what a face looks like, for example, by looking at a bunch of images of faces.

A better analogy is, let’s say I grew up in an isolated mountain town and never left my mountain. I’ve seen a bunch of movies and tv shows where the characters go on rollercoasters, but I’ve never seen a rollercoaster in real life.

And let’s say I’m a big fan of Legos. I’m able to use my understanding of what a rollercoaster looks like to build one out of Legos. That doesn’t mean I’m copying bits and pieces of rollercoasters I’ve seen in movies and tv.

Similarly, this AI system doesn’t store a bunch of images and use those to create a “new image”. Going back to our face-generating AI example, it’s creating a generative model by processing a large number of images of faces and mapping the features to a probability distribution. Then when it tries to make a new face, it spews random data into a function that weighs the random data by the previously-calculated probability distributions. So random data is “massaged” by the probability of what a face looks like, like the probability of an eye being in a certain location or the color of the eye or the shape of the eye.

Once this random image is created, a separate neural network checks to see how well it matches its understanding of what a face looks like. This neural network is also trained on the dataset of a bunch of faces, but it’s not outputting any new images; it’s just saying “yes this random image looks like a face” or “no this random image doesn’t look like a face, and it’s off by this much: x”. This process iterates, helping the generative system get closer and closer until it generates an image that’s good enough.

Both components of this system are using probability distributions for a bunch of variables. They don’t even store any images.

So I really don’t think there should be any legal liability here. If there is, then Bruno Mars should be sued for hearing Michael Jackson’s music and using it for inspiration.



I hope that helps explain things a bit, and it's certainly the way I understand the workings of AI, but it's good to air our views and have a sensible debate about it.

To add ..

Without going into too much boring detail, these AI programs are based around Neural Networks. The more advanced networks don't just take words from the text prompts and search for existing images with matching words in their filenames then composite them. No, they use the prompt text to produce a 'thought' ie. a picture in the 'mind' of the AI as to what to look for, then it goes looking.

Translators now use much the same process. They don't just translate, say, a French phrase to English by comparing it word for word and using a dictionary lookup. They first read the full phrase and form a 'thought' of what is being asked in French then use this 'thought' to compose an English phrase closely matching the 'thought.' This produces much more accurate translations.

Honestly, these AI's are very close now to actual sentience.