AI Art: It is criminal to not post your prompt

The Something Awful Forums > Main > General Bullshit > AI Art: It is criminal to not post your prompt

namlosh: Feb 11, 2014; I name this haircut "The Sad Rhino".

Sorry if this is the wrong thread, but what are some of the best models that can be run locally?
I�ve got it stable diffusion running locally from this git repo:
https://github.com/CompVis/stable-diffusion.git
And it�s generating images but they aren�t that great. I�m not trying to generate stuff as good as whats in this thread, but I am trying to generate the best I can locally for free. I want to see what the limit is for offline generation. I have a beefy machine with a 4090 if that matters.
I�m using the �default� stable-diffusion-v1/model.ckpt file that�s around 7.5gb. Should I be grabbing a different one from somewhere.

Sorry if this is all obvious or the wrong place to ask. Just point me in a direction and I�ll head that way.

# ¿ Mar 25, 2024 17:06

Adbot: ADBOT LOVES YOU

# ¿ May 14, 2024 03:48

namlosh: Feb 11, 2014; I name this haircut "The Sad Rhino".

dumb. posted:

I'm not familiar with that SD ui, but it looks pretty old... can it only do SDv1? If so, you might want to check out a more modern interface that can run SDXL models like Fooocus, AUTOMATIC1111 or ComfyUI.

As for XL models, I'm a fan of RealVisXL, Juggernaut XL, Colossus Project and ZavyChromaXL.

I don't use 1.5 very often anymore, but some 1.5 models I dig are Analog Madness, Photon, Realistic Stock Photo and Epic Photogasm.

Awesome, lots of keywords for me to google. I�m sure that I�m probably using a UI that�s old� it�s command line lol. It just looked like the official one so I started there.
There�s so much crap out there for AI now that it�s hard to search for real stuff.
Is there a good resource for searching available models that are somewhat vetted? I�ve used hugginngface but it has so many models that are just blank initial check-ins. Is it the best resource to see what�s out there?

# ¿ Mar 25, 2024 17:56

namlosh: Feb 11, 2014; I name this haircut "The Sad Rhino".

Sedgr posted:

https://civitai.com/ should be helpful for some models and such.

The Sausages posted:

As far as I know you've managed to install the OG version of Stable Diffusion, I'm kinda impressed that you went there of all places. If you went that way because you want to learn more about how it works I'd recommend this old version as well, it explains what the software is doing.

otoh if you just want to make pics try Stability Matrix, it's an installer/manager for SD versions and models/loras/etc. It lets you browse and search civitai as well, no CLI/python/git needed. It also has a decent built in GUI that runs on a comfyUI backend.

As far as "vetted" models if you stick to the most popular ones on civitai.com they tend to be good. Imo sdxl is consistently better at higher resolutions but with upscaling such as a1111's hires fix it's less crucial. You can also filter checkpoints by "trained models" which helps for originality a lot of merged models are kind of an inbred mess. Definitely recommend browsing civitai with the nsfw filters on if you aren't desensitized to internet filth - heck I feel pretty desensitized but I still leave the filters on when seeing what's new.

Another tip for beginners is to find pics you like on civitai and copy the generation parameters. Prompts aren't so important as having a sampler/steps/cfg combo that works consistently for the art style you want.

Objective Action posted:

I quite like https://github.com/LykosAI/StabilityMatrix for new users as it lets you one click install any of the popular front ends, automatically consolidates the models between them, and has hooks to browse and download models right from the interface.

e: fb, what I get for leaving tabs open

mcbexx posted:

I am still a huge proponent of InvokeAI - https://github.com/invoke-ai/InvokeAI - if you want to give that a twirl. Have been using it since October '22, it's pretty neat, all in all.

Fairly easy install and comprehensive help/troubleshooting pages for when something goes wrong.

They have some tutorials on their YT channel to help you familiarize yourself with it.

This is all amazing and I'm really thankful for the help.

I am a software developer and I do mess with python a lot. And I am looking into learning more about the back-end of these. For example, I'm also messing around with Ollama for gpt type stuff.

I've got the cli stable-diffusion1.0 working along with invoke-ai now thanks to this thread. I was planning to try a bunch of different options out at first... AUTOMATIC1111, ComfyUI, etc... but then I realized I'd have to re-download all of the models...
But it sounds like StabilityMatrix would be able to help with that problem? Maybe I'll install this next.

In the meantime I've also been able to get SDXL working... let's see if I have some of the terms right:

Model: stable-diffusion-xl-base-1-0
Checkpoints: realvisxlV40_v40LightningBakedvae
Refiner (or is this just another checkpoint?): sd_xl_refiner_1.0
ControlNet: A way to add a picture to control the generation of other pictures? (Haven't used one successfully yet)
IP Adapter: same/similar to the above? (Haven't used one successfully yet)
LoRA; zero clue... but I'll figure it out

Actually, I don't want to waste everyone's time answering my questions that I could probably easily look up. I just wanted to jot down some stuff and give a progress report. I wouldn't even be here at all if the thread hadn't helped out.

thanks all!
ps: prompts(both +/-) are hard, lol. but dang this is fun.
I kind of feel like I want to grab sections of literature and feed it in as prompts and just let it run through a book/newspaper overnight lol

e: well dang... StabilityMatrix won't install any of the UIs correctly. It has an unhandled exception that blows up the entire app before I can even copy the stacktrace to fill out a proper bug report.
I guess I'll just wait for another release then. Tried this one: https://github.com/LykosAI/StabilityMatrix/releases/tag/v2.9.2

namlosh fucked around with this message at 02:31 on Mar 27, 2024

# ¿ Mar 27, 2024 01:15

The Something Awful Forums > Main > General Bullshit > AI Art: It is criminal to not post your prompt