anyone using a GPU for AI art?

[Threepwood]

hello all,
as per title, is anyone using thier GPU for generating AI art?
if so, can you guide me to the first rung of the ladder?
I don't know where to start! I want to do this

 

[decto]

I've had a dabble, but have to many other things but is still installed for later.

Watch a few you tubes on installing Stable Diffusion or Automatic 111 (A front end for Stable Diffusion that helps with the prompting)
Install guide here: Automatic 111

Depends on your GPU somewhat, Worked fine on a 3060 12GB at ~ 10 secs for a pair of 512 x 512 images.
That is one thing to keep in mind, generally you make small images which can then be AI upscaled later.

I think stable diffusion XL renders 768 x 768 but needs more GPU resources.

Apparently it will run on a GTX 1650 4GB.... but very slowly. 8GB+ Nvidia is better though there I think you can run AMD under Linux with ROCm. though I haven't done this.

 

[Mr Evil]

I do. I make deliberately bad images for use as birthday cards and such. I use SHARK (Stable Diffusion), which is nice because it's very easy to use - just one executable that does everything. Works on both AMD and nVidia GPUs.

 

[eeii]

I do. I make deliberately bad images for use as birthday cards and such. I use SHARK (Stable Diffusion), which is nice because it's very easy to use - just one executable that does everything. Works on both AMD and nVidia GPUs.

Nice one will give this a go.

 

[Donnie Fisher]

Fooocus is a good place to start I found.

 

[Emlyn_Dewar]

SD Next ftw

 

[Steviepunk]

I've dabbled with ComfyUI, allows for some good workflows and options but does take a lot of setup work (and I'm not sure I've got mine setup correctly).
However, I'd suggest starting with Automatic1111 to get the feel how it works.

In terms of base models:
SD 1.5 - a lot of resources, Lora's, etc available however can have issues with hands, etc if not prompted correctly (seems to rely a lot on negative prompting). Max res: 512x512 before upscaling
SD 2.0 - not tried it, I don't think the community got behind it. I believe max res is still 512x512
SDXL - more matured model and better overall quality of image generated (eg. better handling of hands, etc), good community support. requires more resources that 1.5, however max res is 1024x1024
SD 3 - Very new and the community seems to be turning against it, the publicly released model seems to have problems. In theory should be more efficient than SDXL while producing better images. Max Res 1024x1024

The style of prompting between SD 1.5, XL and 3 vary quite a bit, so a good prompt for SD 1.5 will not necessarily produce a good output in XL (and vice versa)

Not a definitive guide there, but hopefully gives you info to get started.

If anyone happens to know any Checkpoints/Models that are good for creating images of Transformers (Optimus Prime, Megatron, etc), please point me in the right direction - I've found Dall-e and a few other online systems to be reasonable good at them, but Stable Diffusion just doesn't manage it (though not yet explored Lora's for that)