AI

[shawnhcorey]

5 minutes ago, Malkavian Grin said:

I'd say the only difference is using what we as Humans consider "proper interpretation" of things such as how physics will interact with objects and the like.

 

I disagree. Intelligence is more than really good algorithms. Intelligent systems need at least a self-image. This allows the system to think both subjectively and objectively about problems. It makes it a self-referring system. Self-referring systems give a paradigm shift to what is possible. The machine-learning algorithms of this "AI" will always have limitations on how intelligent it can become because it lacks this ability.

[TheFred]

If anything, I think the machine-learning models are probably better than most humans here. For some reason, a lot of artists aren’t also into geometry or physics; I’ve seen a lot of messed-up perspectives, and not everyone can do clouds like Constable. I reckon a lot of people just do that intuitively, in exactly the same way that AI learns the pattern without really understanding it.

I’m not really sure how one defines a self-referential image or proves that it’s necessary for intelligence. At this point, we’re into the philosophy of what it means to be sentient. Given that we’ve no way of proving even that other humans are sentient, I doubt we understand what would be necessary for an AI to be so. At any rate, no, these are just models, but I think humans brains behave a bit more similarly than most people think, building their own models and matching patterns in a very similar way (indeed, neural networks are so named because they’re inspired by the structure of the brain itself).

[shawnhcorey]

23 minutes ago, TheFred said:

I’m not really sure how one defines a self-referential image

 

I mean the program, not anything it might create.

[Malkavian Grin]

@shawnhcorey I think TheFred also meant the program, since we're talking philosophically now. Can you more define this nebulous "self-referential image"? Do you mean a consciousness? Do you mean being self-aware?

[shawnhcorey]

21 minutes ago, Malkavian Grin said:

Do you mean a consciousness? Do you mean being self-aware?

It may be conscious or not. It would be self aware.

Inside any AI is the world model. This is the database of the facts it knows about the world. This lets it think and plan about actions it may want to carry out.

Inside the world model are facts about the AI itself. Collected together, these form the self image. There is nothing special about these facts. The AI can use them like any other fact in its thinking. It can think objectively about itself.

But what is special is when the AI thinks subjectively about itself. A trivial example, say if it has to navigate thru a building. It only knows about the room it is in. But to find out what's in the next room, it can send itself to that room to see. It can use itself to expand its world model. The AI is no longer randomly exploring or following an algorithm. It is planning it own future. This may be trivial to us humans but it's a big step up in thinking ability. It now has the ability to plan what it learns.

[DarkisNotEvil]

This is kind of drifting into philosophy instead of application, so I'd like to talk about things like character portraits- I think they were mentioned earlier. You can use an example generated image as a template for describing a character. You don't have to use it as an image in the game, but sometimes it's good to literally put a face to a name, so you can describe traits you see visually you wouldn't normally describe in order to be more descriptive (beard shape, eye color, nose shape, little tokens or knicknacks), even if the image is never posted.

Edited May 25, 2023 by DarkisNotEvil (see edit history)

[Malkavian Grin]

11 hours ago, DarkisNotEvil said:

character portraits

I have found a pretty decent (and free) software called Dream - by WOMBO that I use for quick portraits for just such a thing! I've been using it in my solo game Castle of Memories when I have a basic idea of a character but nothing very specific (say I know it's a black woman who can tell things about the future, so I type in "fortune teller"). It helps me visualize things relatively quickly and even assign a kind of personality to them instantly just based on how it looks and feels. I'll include two examples.

The left is my character and the right is said fortune teller. It works decently but... don't look too hard or you'll find glaring errors.

Portraits

[TheFred]

Those are really cool!

I've had some success, albeit mixed. I tried making portraits for some of my existing characters.

For example, I have a character who was partially inspired by the picture on the left (which I've struggled to find the original source for but appears to be by someone called MatBudimir who since left DeviantArt where it was posted). Stable Diffusion was able to produce a lot of pretty cool images that look quite similar; here are a few (it also made one where she appears to have laser eyes, but c'est la vie).

 

This cool picture made me want to make a superhero character for a superhero game. Turns out, the AI can generate superhero pictures OK too, though it does tend to try and give people a Superman logo or make them look like Batman...

(It's pretty easy to get it to do different styles, too, like the one on the far right)

It doesn't really like fantasy/nonhuman stuff so much, as I guess it has less reference for them. You have to try quite hard to make it do unusually-coloured skin, for example (bluntly, most of the models seem pretty biased towards white people, but if you ask it will do whatever; getting a red-skinned tiefling or violet-skinned drow is quite hard though and you're probably better off recolourising manually afterwards, though img2img helps).

Here's a tiefling by dashinvaine and some AI attempts (I struggled a lot here as the tiefling character I was trying to create at the time is nervous and sad, and so I had to discount characters who looked too imperious or badass, but I was able to generate other decent tiefling pictures generally):

 

And, that said, if you want to be a fox-person (or, like Master Splinter type character?), that's fine too:

 

Anyway, I shall stop now before I hit some kind of image upload limit or something...

[Malkavian Grin]

@TheFred Woah, how many generations did it take you to get those various results?

Agreed, most of these generators seem to be biased toward white people which is aggravating and unfortunate. I have a project I'm working on slowly, which uses a deck of playing cards. I decided I'd use A.I. to create brand new face cards, with a wide diversity of people. Some worked better than others... It's almost like no one ever painted Black or Indian people during the renaissance...

I have to say I'm impressed with the fox images! Makes me wanna make lizard-folk and Ninja Turtles ❤️ ❤️

[TheFred]

I believe that you can train your own model on just a small number of images, using an existing one as a base (training from scratch requires a huge dataset); people have done things like make specialised models for specific styles (there being some controversy there, where e.g. someone's trained something on a specific artist's work). Honestly I see what we have now almost as a proof-of-concept (albeit a very impressive one!) and I think it's just important to be aware of the biases and limitations (more so with something like ChatGPT to be honest; it can do some amazing things, but when was it ever a good idea to take at face value what some random, person or otherwise, told you on the internet? Or even just in real life?).

I've definitely picked some of the better images there. They all have little artefacts and oddities especially if you zoom in, but it's not hard to get stuff like that if you generate a bunch. And, you can use a base image for it to make something similar to, and do inpainting where you get it to regenerate a section, which can be very useful. I think the harder thing is not getting a "good" image but a specific one. It struggles a bit with complex compositions ("jackal-headed character" is fine but "jackal-headed character carrying scrolls" it really doesn't seem to like), though inpainting can help you with this, or doodling a base image. The models also have associations, so if you ask for one thing you might get another thing which is commonly seen with it. A prime example in my experience is colours. If you ask for someone to be wearing a particular colour, you often get that colour for their hair as well, and sometimes you get the hair but not the clothing - or vice versa.

It weirdly seems to be really good at doing head-and-shoulders portraits (which you can just ask for explicitly) but not so good at doing faces as part of full-body shots. You'd think you could just zoom in somehow and inpaint that bit but I haven't had huge amounts of success.

I've been using a lot of Dreamshaper for the model which is more fantasy-art trained and a model I made by merging it with another one (I find it's usually just a bit too fantastical, with a bit of an anime lean, and gives things that look like they're off video games - they look good, really impressive, but for some things I like the results better if toned down slightly).

[Malkavian Grin]

Ahh, it sounds like you've installed your own model somewhere then? I don't have the resources to be able to do that. Sounds like a really neat experiment though!

Some programs I've found do better with different modes of explanation too. For instance DALLe prefers long-ass sentences separated by commas, but WOMBO's Dream seems to like shorter, more poignant statements separated by +'s.

[TheFred]

I’m just running it on my PC; you need a pretty decent GPU, but nothing special (mine is a 4GB one which is just good enough). The models themselves are not the hard part, they’re big-ish, like ~4GB, but that’s fine. I’ve actually been using Easy Diffusion which has an option to use the CPU instead but it’s meant to be very slow.

[Neopopulas]

I've found some pretty good ones that let you iterate over and over to get what you want. Some are better than others of course but i use both https://playgroundai.com/ and https://creator.nightcafe.studio/studio from time to time and they can get some pretty good results. You do kinda have to get used to how they work and it takes practice but it gets some surprisingly good results these days.

[Malkavian Grin]

17 hours ago, TheFred said:

I’m just running it on my PC; you need a pretty decent GPU, but nothing special (mine is a 4GB one which is just good enough). The models themselves are not the hard part, they’re big-ish, like ~4GB, but that’s fine. I’ve actually been using Easy Diffusion which has an option to use the CPU instead but it’s meant to be very slow.

I have an old PC so I can't run stuff like that. Still neat though!

11 hours ago, Neopopulas said:

I've found some pretty good ones that let you iterate over and over to get what you want. Some are better than others of course but i use both https://playgroundai.com/ and https://creator.nightcafe.studio/studio from time to time and they can get some pretty good results. You do kinda have to get used to how they work and it takes practice but it gets some surprisingly good results these days.

I'll give these a try soonish and report back with my findings 😉

[fabulist]

I used nightcafe to make the images in here.

The prompt was 'a marshland at the bottom of a canyon'.