“Album cover for new album by The Beatles”
Been messing around with D3 today. Super impressive that they seem to have cracked positional awareness. I was able to make a few images that were impossible in D2 because they had too many components.
Can you recommend a minimally complicated path towards getting a client up and running? I looked around awhile back and the space is so flooded with enthusiastic nerd discussion deep in the programming weeds that I couldn’t figure out a starting point.
The easiest would be to download NMKD with the included SD 1.5 model. Then look up any additional models you might like from Civitai searching for “photorealism” or whatever catches your interest. Pop them into the appropriate folder and you’re off to the races. As I recall, you recently bought a new PC so you should be fine, spec-wise.
Pre-emptive note that there’s a LOT of “pretty girls” style models on Civitai though I think you need to opt in to seeing the adult stuff. But there’s also a lot of perfectly respectable SD models that you could show your grandma.
I know that the NMKD dev is working on SDXL support but I don’t know if the latest client on itch supports it. That might require a jump to using Automatic1111 as your client which is a bit more complicated but not terrible (hey, if I can do it…). NMKD is an easy way to be up and running in very short time and with minimal tech input; if you can click an installer, you can set up NMKD.
SDXL is the latest/greatest stable diffusion model, right? I’d definitely like to try that, but I’ll start with NMKD and go from there. Thanks!
Yeah SDXL is the latest.
I don’t think NMKD 1.11 supports it; at least it won’t even show an XL model on the list. 1.12 says it does but that’s a beta off of github and a little more work to set up. I was hoping 1.12 would be live on itch by now.
Worst comes to worst, you decide that 1.5 sucks and take the plunge into using A1111 for SDXL 
Followup: do these models all act independently of each other or do they layer on top of the base model to weight results in various directions?
Models are independent. They were trained using the base 1.5 model with an eye for greater realism or anime or whatever the creator was looking for. You use one model at a time.
There’s also something called Loras which act as an overlay. So you’d load up your photorealism model and then add a Lora for specific attention, such as better detailed firearms or a comic book artstyle. As far as I know, there’s no technical limit to how many Loras you can use at once but more of a practical one as they try to influence different stuff.
Neat! AI art stuff was one thing I had in mind with the new PC so I’ve been meaning to jump in for a while.
Can you point me at whatever instructions you used to get A11111111111 running? I’m already seeing a dozen different tutorials.
One cool thing about Loras is that they’re relatively easy to train yourself. Training a whole new model (not from scratch but using SD 1.5, etc as a base) is a whole undertaking that most people wouldn’t want to try unless they have an RTX 4090 or buy GPU farm time online. But a Lora is lightweight since it’s more like an influence or nudge and can be done yourself. So, if you were really unhappy with how SD was rendering squirrels, you could round up 50 pictures of squirrels, train a Lora and have that extra “Ok, this is what a squirrel looks like” when prompting.
With the GPU being the biggest concern. He bought a new PC, but did he put in a $500 or more Nvida card with 8 GB of RAM?
I believe so but the client has settings for low VRAM cards as well. Just gotta be patient and not try to natively render large images.
If you want to try out models, go to Playground and try various “filters” (as they call them) on your prompts. Some of their “filters” are just extra prompt text tacked onto the end of your prompt, but others are full models. They have different “filters” for SD15 and SDXL (and Playground v1).
It is nicer to use someone else’s H100.
I have a 4070ti, which should be good enough for government work.
They all have their benefits. I still keep my Midjourney account but it’s nice to be able to render as many images as I want of whatever I want without worrying about credits or time-outs or a service deciding that “Elon Musk eating a gyro in one bite” is no longer an acceptable prompt. I was making some player character images for a superhero tabletop RPG and probably spent 100+ infill renders trying to get some stubborn details correct and thought nothing of it. But using someone else’s system means you don’t need to worry about your own hardware or software or chasing the newest updates, etc which is nice as well.
In this case, the GPU in question is an RTX 4070Ti so it should have no issues. It’ll crank out 1.5’s native 512x512 images with a swiftness and SDXL’s 1024x1024 at an acceptable rate.
Yeah, that’s one of the fastest.
I brought that up because not everyone buying/building a PC gets a premium card. For instance, not being a gamer, my PC runs on whatever video is built into my CPU. It wasn’t an issue for me until SD came along and made me want more. (In the past when I got spendy on video cards it was for NTSC inputs and video capture. I bought two different ATI All-In-Wonders, if anybody remembers those.)
@Jophiel, I think I edited this into a post after you’d already responded to it.
Thanks again for the handholding. I’m normally fine with a little computer DIY, but I do have trouble getting started with navigating stuff like this, where nothing is really centralized and every website feels iffy as fuck.
Your best bet for SD is the main reddit group.
I started there, it’s just got too much signal:noise ratio in terms of getting started from zero knowledge.
Jophiel sent me a link to a helpful tutorial with an offensive name, so I’ve got a path forward.