You can run SDXL perfectly fine on a 3060 12gb. That's like $300, hardly a supercomputer.

@randomlettershdhtyxod yeah that doesn't fit in my case, nor do I have a power supply for it, and $300 is a lot of money in my country. Idiot

@MorganFreeman These guys provide these models for free and you complain because you can't afford what is needed? Before you start insulting people because YOU can't afford anything, let's see you try training a model and upload it here to contribute to the community. Oh, that's right, you can't because you don't have a GPU that could handle it. Be happy these models are free and if you can't use it, move on. It's not their problem YOU can't use it because of your hardware.

Look up nmkd stable diffussion gui on itch.io. Update the program. Put the lora and checkpoint files in the right folders. It might be slow but if you run the image at 512x512 on a old computer with 24 steps. The image might take 8-19 hours on low end hardware from the 2000's if you have a semi good computer you might beable to render 720p images in 4 hours time. Otherwise, you need a nvidia computer to render stuff fast.

@MorganFreeman Very true, $300 is a lie, you need to buy matching mobo and CPU not forgetting DDR5 one, as I have DDR4. In the end $300 is a lie, in reality a lot more money needed to "upgrade" (more like replace).

It's interesting, cause here I thought everyone would want to keep these things "licensed" like this to keep it free and open to the public. I didn't think that if said model is licensed in this way that it then can't be used on generation software for people that don't have the income/situation to utilize this hardware. It's almost as if SD purposefully made sdxl more hardware intensive just so they could start getting people to buy into all the generation services out there cause they can't run it locally as easily as 1.5. However, that's just my tinfoil hat theory. I can see both sides to it now, but still prefer it to the public, for even if you can't generate on it now... it's not going anywhere, so if anything it's an incentive to get the money/situation figured out and then you have this as reward, plus all the other amazing models these awesome creators have given to us.

@trytensor sounds like a lack of researching on the users end. Using ai is no different than someone buying a integrated GPU vs a dedicated GPU.

If your gonna buy a low end computer. You shouldn't expect any good results or fast computing .

@FoxMccloud2022 Just little correction there, not gonna buy, but had already bought, long time ago.

@trytensor To correct the corrected. Just because, the computer you have may have ben modern doesn't make it usable with most ai programs. Unless you have sufficient ram and vram you wont get far with ai generation. Even with my intel pentium 2 with invoke ai I can run 512p 22 step in around 8 hours. So theoretically if your computer has 8 gb ram and vram unlike my computer with 4gb ram and 256 mb of vram. Your computer should beable to create art in a matter of minutes.

@FoxMccloud2022 Yeah, my point is neither I bought low-end gizmos due to lack of research at 2016 DDR4 is the king, nor I'm happy somebody just saying "buy 3060 now it's cheap only few bucks not supercomputer" if you get what inb4 addressed.

@trytensor If it helps. Why not get a crucible 20 gb ram card for like $20 and a usb with $50 -80 gb of ram for $40. So you could use the usb for ready boosting to help act as vram. I know its not as good as upgrading hardware. But, adding more ram and ready boosting might help improve ai generation. If my old as heck computer can do art. I'm sure all you need is something alot less cost effective. Since, it's not like you need a gaming machine. What resolution are you trying to create art with?

@FoxMccloud2022 Well, I'm mostly only trying few models here and there, more curious towards motion and img2vid like stuffs. Getting 1024x768 already made me happy, if rarely 1344x768, but for 1.5 stuff usually 768x512 is enough, since there is tile upscaler. The dimension here is w x h.

Try it with Euler / Euler Ancestral. It seems to fix the problem, for some reason. Though you also need the recommended negative prompt.