Stable Diffusion AI Image Generator

OnlyOneKenobi

OnlyOneKenobi

Expert Member
Joined
Dec 6, 2013
Messages
3,837
Reaction score
5,178
This is sort of a continuation of the DALL-E2\AI image generator thread, but I figured SD deserves its own thread since it's open source and pretty much the only one of these tools which can be installed, modded and run offline without subscriptions, and so on.

I figured those who are interested can talk about the tool, prompting, how to install and mod it, showcase their generated pics, etc in this space.

Anyhoo. I've been playing around with some of its vast capabilities, and one worth mentioning today is its image-to-image generation which allows one to, among other things, take things like old school video game sprites and upscale \ redraw \ HD-ify them.

One example is the good ol' original Doom guy as seen here :

grin.jpg

And here's his 2022 AI redrawn \ upscaled version. The AI gave him glowing green eyes for some reason, but I liked it somewhat and decided to keep it.

Doom Guy.jpg
 
Fulcrum29 said:
How well does this do with AMD GPUs?
Click to expand...
Not sure. I have an RTX3060, so I haven't really looked out for SD running on PCs with AMD GPUs.

There are versions of SD that actually run on M1 Macs, so maybe there is a version that will work with AMD.
 
OnlyOneKenobi said:
Not sure. I have an RTX3060, so I haven't really looked out for SD running on PCs with AMD GPUs.

There are versions of SD that actually run on M1 Macs, so maybe there is a version that will work with AMD.
Click to expand...

Through reading, it seems that AMD GPUs are doable, as long as the minimum ram requirement is met. I think I have everything which I need to run SD on AMD, but I haven’t had the time to set it up. Just asking because I don’t want to go through the steps only to end up with something which is borked.

Basically I want to reappropriate a server to do AI image generation, and run input queues through it which I can sample by day end or the next day. I have a Vega GPU, old, but it should work. No Tensor Cores, but that I can rig up at a later time with an RTX capable GPU. At this I want to bench what I am attempting to achieve.
 
So here's another example of SD's image to image capabilities, generally, most images will require some kind of cleanup or minor tweaks, but it's quite amazing what can be done if you guide the software with a combination of source images and descriptions. I took a sprite from the old videogame, Redneck Rampage, ran it through a few SD batches, found a good match for the original look and upscaled it. The hands are a bit wonky, but considering how much work it does just out the box, it's quite amazing.

RRampageUpscale.jpg
 
I was wondering how long it would take for people to catch on to this.

/aco/ had a few threads up 2 or so months ago showing people how to set up.
 
A bit more info for anyone settup up...

 
Works quite well taking a very basic sketch and letting it do the design work.

Prompt text: desktop lamp, glass dome, futuristic, light, base, warm glow, geometric
Prompt image:
ty.png

Output:
fr.png
 
saor said:
Jeez this looks like a lot of fun.


edit:
Thanks, got it up and running.
Click to expand...
The img2img stuff is what I want to learn next.
People have been able to clean up images, and even replace parts of an image.

Very useful for clearing up text2img results too.

Definitely need to watch.

What GPU you using? Even an old 1060 6GB works very well.
 
Vorastra said:
The img2img stuff is what I want to learn next.
People have been able to clean up images, and even replace parts of an image.
Click to expand...
I'm finding even with verrrry basic inputs it creates useful output. I mean if I had no design inclination other than a crude notion of what I was after, I could feed in the rough drawing, some rext prompts and end up with results that could be feasible options to fabricate. Be cool to see this generated in 3d.

rfrf.png
Vorastra said:
Very useful for clearing up text2img results too.

Definitely need to watch.

What GPU you using? Even an old 1060 6GB works very well.
Click to expand...
RTX3060
Yeah I'm also curious to go down the modifyingimages rabbit hole. Will probably also start building a model of some family and myself at some point and see how that turns out.
 
Also cool how it preserves color and general geometric layout of the input. In the dragon one it doesn't seem to see the input as a face but rather as a reference for how the image should look stylistically.

Input: This image + some text about dragons etc.
234426.png

Output:
tmp4otz1e2u.png
tmpmmwa10b0.png

I like what it did here...
dede.png
 
The potential of this tool is massive. I've been playing around with its deforum plugin (video) and it's just jaw dropping.
 
You must log in or register to reply here.
Back
Top
Sign up to the MyBroadband newsletter