Fun with Stable Diffusion

[JEDI]

 

[JEDI]

Steve Jobs (Rip 2011, cancer)

 

[Red Squirrel]

This needs to be a movie.

 

[JEDI]

Happy 2024!

 

[Ajay]

Elvis on a Harley eating a banana.

 

[JEDI]

Behold.. the Santa Cardinal

 

[Shmee]

Birb!

 

[IronWing]

I was trying to make an image of people swapping fluorescent minerals and this missed the mark but it is also weirdly awesome.

 

[lxskllr]

Art deco meets cyberpunk.

 

[Red Squirrel]

It's fun to play with the different styles to see the results. A couple Renaissance style ones.



Worker installing cables



circuit board assembly line



DMS100 technician installing line card. (that's super industry specific so was not expecting anything remotely close lol)


Quantum computer


Fusion reactor

 

[snoopy7548]

Alien zombie attacking a bowl of spaghetti

 

[snoopy7548]

It's amazing how far this has come in realism and quality.

Here was my "hot chick with pointy elbows" about a year ago:





Now here it is:

 

[Red Squirrel]

Elbows are not pointy enough.

 

[Red Squirrel]

"darth vader choke hold" in misc-dreamscape.

This looks like it could be a cover for a Star Wars valentine's day special.

 

[Red Squirrel]

Couple randoms using advertisement filter. Filenames are what I put as prompt. Mish mashing products company/franchise name to see what it comes up with.

I like the "Super Microsoft Entertainment System" it basically just did Xbox running Windows lol.

 

[IronWing]

Looks like they want an account now.

 

[lxskllr]

Well, that's no fun.

 

[lxskllr]

AI is stuff I'd like to do on my own computer. I'm not really into subscription services. I wonder how hard it would be to do on your own? Seems like a company could sell an AI starter kit with a bank of drives in a box for data and some software(using an open standard of course) for generating media. It could then be updated to fine tune it for each individual. Would you need a gfx card for speedy calcs?

 

[igor_kavinski]

Would you need a gfx card for speedy calcs?

Yeah. My friend is doing it on his 3060 12GB.

And yes, you do need a 3D card or at minimum, an AMD iGPU with 32 GB RAM so you can allocate more RAM to the AMD GPU.

He's using this: https://github.com/LykosAI/StabilityMatrix

For AMD APU:

https://www.reddit.com/r/StableDiffusion/comments/12of1o5

This one works with 5600G: https://github.com/lshqqytiger/stable-diffusion-webui-directml

 

[Pohemi]

Also, you need some storage space for the model sets (buddy does it offline and says his current model sets take up around 300GB so far, and says he's constantly DLing new sets).

Edit: my buddy is running a Ryzen 5900X and RX 7900 XTX, and says some (renders? idk what you call em) pics still take a while to construct, like 30 minutes. With a beast of a GPU like that, it must eat up a lot of computing power to still take a while like that.

 

[Red Squirrel]

I want to look into self hosted Stable Diffusion at some point, it sounds fun to be able to play with that locally. Would be fun to train it from scratch too and I presume you can also download pretrained data.

I could finally find a purpose for my mining rig, just need to find GPUs...

Something like ChatGPT could be fun too. Maybe create separate small instances that are trained on specific data, like have one trained on far right stuff and another trained on far left, and have them talk to each other.

 

[igor_kavinski]

Something like ChatGPT could be fun too. Maybe create separate small instances that are trained on specific data, like have one trained on far right stuff and another trained on far left, and have them talk to each other.

Oh you are kinky

 

[Red Squirrel]

Could go even further, like have one trained on how Intel is better, and another AMD. Also need at least one trained only on Linus Torvalds rants.

 

[daerragh]

You cannot start generating AI images more easily than with Fooocus software. The installation and running is fully automatic. It needs a mediocre GPU (I use GTX 1660 Ti 6GB Turing) and generation (fully local, on my GPU, no internet and no censorship) of a 1024x1024 image takes only 1 to 3 minutes (just seconds on better RTX cards).

Everyone should just start with Fooocus (weighs 20GB with 2 built-in models) and, when you learn the basics, use other GUIs like ComfyUI and other models like dreamshaperXL_v2Turbo (free at civitai.com, weighs only 6GB).

Start here: https://github.com/lllyasviel/Fooocus
Download, unpack, run updater .bat file, and finally run.bat to get the whole thing going. It's super easy and yields in very satisfying images.

Just a few hours of learning and trying various GUIs and models (they are called checkpoints) and you get this:

 

[Red Squirrel]

Been meaning to look into self hosted AI, it sounds fun to play with. Might be an excuse to revive my mining rig as it's already setup to handle up to 8 big GPUs. I don't know if I would go that far as to populate it though. That's a lot of money.