Stability AI

[sfoster]

I downloaded a local copy of stability ai (<--- that means link to free download) a text to image AI program.
Made this it's own thread bc it's free to run and everyone can download and use it, not a closed beta or anything.

Kind of addictive atually lol, my system requirements are 576x576 for image dimensions and it takes 37 seconds an image
Here are some of my favorite images from playing around all day yesterday.. starting with my avatar.

Satan in a Christmas Sweater

satan.png

drive.google.com

Wizard

20220827002550_2745022183.png

drive.google.com

Turns himself into a dragon

20220827235941_672055017.png

drive.google.com

And attacks his magic school / lighthouse

20220826235046_3459050376.png

drive.google.com

Destroyer of cities, he establishes his dominance

20220828025148_4294438431.png

drive.google.com

Edit: forgot to add.. while it claims to do nudity and violence, it really sucks at both.
it also completely failed to create a centaur for me. really sucked at mermaids too.

The coolest potential for this technology is to instantly generat complex 3d models, which you can then import into other programs.
Like if you want a bigfoot in your movie, the next generation of this software wil be able to just create a bigfoot for you in 3d that you can import with a skeleton and instantly animate, virtually no work or cost required.

Or like for nate, in a few years this will be able to instantly create a cat INCLUDING built in animations that are modeled exactly after a real cat, just created out of thin air by AI. coming soon.

 

[Nate North]

Some of these are pretty good. Thanks for the link, I try to stay up to date on everything, but I missed this one somehow.

There's quite a few groups working on the 3d model gen stuff right now, and some really basic stuff is already available. You're right though, it's only a matter of a few years before we can begin using text driven AI models in game. A new autorigger came out this last week, which is one component, and there are tutorials for manually converting AI output to geometry. Right now it's more of a novelty than anything useful, but that will definitely change soon.

What people have already discovered though, is that you can feed the AI material textures, which are just square 2d images, and it really does well creating new ones. Repeating brick tiles, rock surfaces, etc.

So it's a great time to start into 3d, because there is about to be an unprecedented explosion in potential for that format. A single user should be able to equivocate the output of large teams from 5 years ago, in just a few years. By some metrics that's already happened with UE5, but once assets and animations can be created procedurally at AAA quality, it's really a new era for 3d design of all types.

The new avatar is great!

 

[sfoster]

Some of these are pretty good. Thanks for the link, I try to stay up to date on everything, but I missed this one somehow.

There's quite a few groups working on the 3d model gen stuff right now, and some really basic stuff is already available. You're right though, it's only a matter of a few years before we can begin using text driven AI models in game. A new autorigger came out this last week, which is one component, and there are tutorials for manually converting AI output to geometry. Right now it's more of a novelty than anything useful, but that will definitely change soon.

What people have already discovered though, is that you can feed the AI material textures, which are just square 2d images, and it really does well creating new ones. Repeating brick tiles, rock surfaces, etc.

So it's a great time to start into 3d, because there is about to be an unprecedented explosion in potential for that format. A single user should be able to equivocate the output of large teams from 5 years ago, in just a few years. By some metrics that's already happened with UE5, but once assets and animations can be created procedurally at AAA quality, it's really a new era for 3d design of all types.

The new avatar is great!

Yeah amazing time for game designers too imo bc, as a software guy, the 3d models and graphics were always the barrier for creating a video game. so cool that you wont *need* a graphic artist anymore.

these text to images programs kinda feel like gambling... lol. you sit there waiting for the slot machine to spin, and then hope for cherries.
and I just keep pressing the spin button hoping i'm gonna find something substantial lol

 

[Nate North]

Yeah amazing time for game designers too imo bc, as a software guy, the 3d models and graphics were always the barrier for creating a video game. so cool that you wont *need* a graphic artist anymore.

these text to images programs kinda feel like gambling... lol. you sit there waiting for the slot machine to spin, and then hope for cherries.
and I just keep pressing the spin button hoping i'm gonna find something substantial lol

I think it's fun playing slot machine with art. In fact, it's kind of addictive, knowing that each time you spin the wheel you might get something amazing. I've had a great time with it, but really the payout ratio needs to be higher before I can consider it serious work time.

For certain types of painted landscapes and portraits, Artbreeder did a pretty good job, but for now there are serious limitations with all of these that mean we still have a little ways to go before it's truly a graphics genie.

My plan so far is to lean into the stylistic capabilities of advancing AI, and keep the core image design process manual. I think it could be amazing to take cells like that Black Rock Canyon reel, and just have them redone as high detail oil paintings, or modern digital art, with the consistency and fluid cinematography intact.

 

[sfoster]

My plan so far is to lean into the stylistic capabilities of advancing AI, and keep the core image design process manual. I think it could be amazing to take cells like that Black Rock Canyon reel, and just have them redone as high detail oil paintings, or modern digital art, with the consistency and fluid cinematography intact.

yes yes very nice, i was thinking more like, put in my favorite movie and have all the actors naked

 

[Nate North]

So what happened on day 2?

are you still throwing ideas at it? Seemed like you got some cool results on your first day.

 

[sfoster]

Yeah i can share a few more

For some reason this evokes feelings in me, idk why but this image does something

20220827234815_1606392999.png

drive.google.com

Dragon on the moon

20220828214523_4007773314.png

drive.google.com

20220828193213_2093311612.png

drive.google.com

Business Wizard

20220828125558_3878851735.png

drive.google.com

Wizard casting a Remote Viewing spell

20220828000246_36005843.png

drive.google.com

 

[Nate North]

Cool!

Some of these are pretty awesome.

 

[James Rogers]

I had some time to play with this program today. Here's one that I came up with.

I call it, MMA fighter Cris Cyborg as painted by Swiss artist H.G. Giger

Oh, and I was able to do a nude woman. I did several but only one turned out well. The others looked seriously deformed.

 

[sfoster]

I had some time to play with this program today. Here's one that I came up with.

I call it, MMA fighter Cris Cyborg as painted by Swiss artist H.G. Giger

Oh, and I was able to do a nude woman. I did several but only one turned out well. The others looked seriously deformed.

looks cool!

I've found that it does breasts pretty well, topless men and women both work out.
but genitals are usually groteque lol

 

[James Rogers]

I had one where the lovely lady had her breasts where her arms should be. I could not tell exactly where her arms came out and I figure I'm probably better off not knowing.

One interesting thing happened. I tried for a rendering of Cris Cyborg fighting Woody Allen and for some unknown reason, Cris was fighting him topless. I mean she was topless, and it wasn't a CGI rendering. Woody's face was certainly CGI looking but not Cris.

One more thing. After 10 renderings, the renderer would render the image then automatically render another image that had nothing to do with anything I described at all. Then the next render after that it would automatically render two extra images that had nothing to do with anything, then 3 extra image... I couldn't figure it out so I deleted the program then put a fresh copy in it's place. It worked for for 10 renders then started rendering extra pictures again.. Almost the way a trial copy of a program will only work for so long then either water mark or do something else to motivate you to buy.. I think it might be because my graphics computer is not on the Internet.

 

[sfoster]

I had one where the lovely lady had her breasts where her arms should be. I could not tell exactly where her arms came out and I figure I'm probably better off not knowing.

One interesting thing happened. I tried for a rendering of Cris Cyborg fighting Woody Allen and for some unknown reason, Cris was fighting him topless. I mean she was topless, and it wasn't a CGI rendering. Woody's face was certainly CGI looking but not Cris.

One more thing. After 10 renderings, the renderer would render the image then automatically render another image that had nothing to do with anything I described at all. Then the next render after that it would automatically render two extra images that had nothing to do with anything, then 3 extra image... I couldn't figure it out so I deleted the program then put a fresh copy in it's place. It worked for for 10 renders then started rendering extra pictures again.. Almost the way a trial copy of a program will only work for so long then either water mark or do something else to motivate you to buy.. I think it might be because my graphics computer is not on the Internet.

you probably have empty new line characters in there.
it generates random images if you just press enter and dont have any text.

i've gotten a few random topless pictures too lol, it's funny like that.
and sometimes it will mix two concepts in an extremely bizarre fashion

once i asked for dragon flying over a lighthouse and i got a lighthouse that had dragon wings.

 

[James Rogers]

About the random images, I don't know. All things being the same, the program just starts doing it..

It's as much fun as you said, but I don't feel there's enough control. Next I'm going to try giving some really detailed descriptions to see what happens...

I don't see too much quality difference between 50 interations and 100 or 200. The program just seems to make different pictures.- and I don't know what the "samples per prompt" means. I've used numbers as high as 100 and don't see any difference....

 

[sfoster]

About the random images, I don't know. All things being the same, the program just starts doing it..

When you enter in a new prompt, try ctrl+a to select all and then press delete. should clear out all the empty space for you.

and totally agree about the lack of control, sometimes i have to dozens of tries just to get one image i wanted.
You can see examples of prompts and search for images other people created here

https://lexica.art/

for example is here is one prompt that i used
a hyperrealistic distant portrait of dragon on the moon, at a sci-fi base station, fantasy, digital painting, dynamic lighting, beautiful, trending on artstation, poster, volumetric lighting, very detailed, 4 k, award winning

sometimes it would give me a sci-fi base station and sometimes a dragon on the moon but i never really got both.
most of those keywords are stuff i coped from the lexica site

this is a great 1st generation tool. gonna be a lot better 2nd generation.

 

[James Rogers]

I'll try the ctr+a delete to see if that fixes it.. Thanks, man!

Yeah, I tried to make a caveman fighting a tiger. For 6 or 7 tries the caveman never even appeared, then when he finally did, he had 6 arms.

Here's Cris Cyborg gonna fight Woody Allen

and here's a kind of scary vampire

One line descriptions for both of these.

 

[sfoster]

I'll try the ctr+a delete to see if that fixes it.. Thanks, man!

Yeah, I tried to make a caveman fighting a tiger. For 6 or 7 tries the caveman never even appeared, then when he finally did, he had 6 arms.

One line descriptions for both of these.

it's not amazing at having two different objects interact, like a cat eating a taco.. usually the taco is not in the cats mouth.

anyway here were my attempts

cavemen.png

drive.google.com

at first i was getting a lot of cavemen / tiger anthropomorphic hybrids, and a caveman with a tiger head, and so i had to explicitly give the caveman a face and that really turned things around for me.

Still not good at the "fighting" part but at least i got something close -- the key is to be very specific, give them both actions, you can make one of them bleeding, etc but you have to be specific in describing what the fight looks like

a hyperrealistic distant portrait of a caveman and a tiger, caveman with a club, caveman with face like ron perlman, snarling tiger leaping with claws, caveman fighting tiger, fantasy, unreal engine, dynamic lighting, beautiful, trending on artstation, poster, volumetric lighting, very detailed, 4 k, award winning

 

[James Rogers]

Yes. I see what you mean.
I just got home from work... Time to see what I can come up with!

 

[James Rogers]

Sean, you were right about blank lines causing the program to render random images. Problem solved!
I'll keep working with the program to see if it can be tamed at all. For the sake of fun, it's a great program, but I've you want something specific, it's pretty frustrating. I suppose the most frustrating thing is not knowing how to "talk" to it or knowing how it thinks. I mean, you can say "car parked on the side of the street" then render it exactly the same 10 times and you will get 10 different images. Why? It's the same description, same random seed. Anyway, here are 4 that I did which turned out similar to what I was trying for.

 

[James Rogers]

I discovered that when the seed is set to -1, the program generates a different seed for each render which means each render will be different looking than the others. If you specify a specific seed such as 238476879, you get the same picture each time you render, assuming the description is the same for each render.

 

[sfoster]

I discovered that when the seed is set to -1, the program generates a different seed for each render which means each render will be different looking than the others. If you specify a specific seed such as 238476879, you get the same picture each time you render, assuming the description is the same for each render.

yes and everytime you generate an image it creates a text file that includes the seed number, so that you can recreate/reword that image.
although its not worked perfectly for me on this app.

if its furstrating also keep in mind that this is the free open source version of earlier software
there is a paid, enterprise version that is called dall-e thats like 15 cents a prompt. its in closed beta right now though