I just had a belated shower thought. How was the guy who made the Taylor Swift images able to find an AI model that wasn't locked down? I know ChatGPT censors NSFW stuff in addition unapproved political opinions. I'm really curious now, because I see all kinds of potential for (nonsexual) mischief. Does anyone know how I can get such a model and how to go about using it? I can think of much better uses for it than making porn of a 30+ celebrity.
Comments (25)
sorted by:
Step 1: Make sure you have a recent generation Geforce video card (the bare minimum is a 6 GB 1060, if you want to get serious about it and make high quality stuff then at least a 30XX series)
Step 2: Install this: https://github.com/AUTOMATIC1111/stable-diffusion-webui
Step 3: Grab a base model that looks interesting from here https://civitai.com/models
Step 4: Start prompting. Good luck!
In addition to this, LORA or lycoris models can add an additional layer on top of the base model. These are the kinds of models you can train yourself pretty easily, IE, based on a celebrity with a small data set of photos.
An older guide and example, but solid enough to give an idea: https://www.youtube.com/watch?v=70H03cv57-o
The guy may have some more recent examples of newer options available, but the LORA stuff I was mentioning was exceedingly solid even 6 months ago.
What about other brands? I have a AMD Radeon RX 6650XT with 8 GB. Are non-Geforce cards any good for this?
AMD cards simply don't work worth a damn for AI generation (as of the last time I read about it, which was last year when this stuff first started coming out): https://www.tomshardware.com/pc-components/gpus/stable-diffusion-benchmarks
I have actually done generation on a 6650XT before. Its mediocre at best, and there are a lot of caveats. Do you daily drive Linux and are willing to do debugging? If so, it's possible to get it working at about half the speed of an equivalent Nvidia card. Otherwise, don't bother.
If you really want to torture yourself with AI on a cheap, janky card, a P40 is a much more rewarding card for a smaller headache.
If you don't want to deal with a bunch of bullshit though, get a Nvidia 3070 or better. I say this as someone who owns mostly amd cards.
I just picked up a 4070 TI Super.
Yeah, that will be 10 times faster than the 6650XT and way less of a headache to set up.
I think there's also open source variants of ChatGPT floating around that have wildly varying capabilities. You could potentially operate these on a private server then upload it that way. If there's one thing I will give the skynet larp crowd is that the voice mimickery is pretty disturbing and that's actually usable. Not been very impressed with a lot of the art and text chat stuff though.
So, just like with the push for the banning of 'assault weapons' it's the everyday, law-abiding people who are going to get punished, not the guy building his own fork and training it from scratch.
Now that there's been a 'controversy' over the use of one of the dncs favorite props- I think social needs to be flooded with revolting, trashy and gross depictions of Swift.
They want to argue for censorship based on sex, which is easy mode for them, it will be way more fun to force it to be argued on fair use grounds and making gross depictions of her
the power of boners always finds a way.
Also, will everyone please just stop using ChatGPT as their fucking gold standard?
Not only is it a poorly competitive AI model, you're just feeding into OpenAI's dominance that is hardly deserved. You can do a LOT better than that, many other people already are.
I feel like we need to have a big-ass 101 discussion here on how this stuff works since so many people don't seem to have the faintest clue how any of this shit works or what's what when it comes to AI options, setup, etc etc.
Please feel free to enlighten us then. If someone points me to something that helps boost my coding efficiency more than ChatGPT does I would be more than happy to use it.
Seeing as no one has been able to do that yet, pardon me while I assume people are not being 100% truthful when they say there's stuff out there that is way better.
You're right, I should be pointing towards specific examples. I might have to put together an actual thread/topic post covering at least a bit of 101 stuff. I don't know why I've always been a little vague in some of my responses on these topics. I mean besides how I'm pretty disorganized.
Anyway, right now ChatGPT is your best bet for generating code and programming solutions, I'll fully fess up to that. But this thread was more focused on image generation, and my comment was pointing to broader AI generated "stuff".
For text generation in general, I can point to a few specific LLM's that I've found a bit solid:
OpenHermes (and hybrid variants), Mistral, and Mixtral (this one is a potential upcoming contender to ChatGPT, but it's still not fully up to snuff on the code-gen). I've heard solid things about Goliath 120-b, Tess, and Nous-Capybara-34B-GGUF, but I cannot verify myself since those were out of my hardware range. There's a lot of solid 7-B and 13-B options, 30-B and up is where it's tougher to find solid options because a lot are sort of "Frankenstein" hybrids. Mixtral is its own thing though, and it's backed with some professional support. You can find most of these just through some basic searches.
A few people I'd mention that are worth following for more info:
https://huggingface.co/TheBloke/
https://www.reddit.com/user/WolframRavenwolf/submitted/
(The former has been a reliable go-to for spitting out quantifized models that are ready to go, the latter is someone who's been at least a semi-useful source for info on potential model leads)
As for front-end and backend for self-hosting, there's a lot of options. There's of course:
https://github.com/oobabooga/text-generation-webui
https://github.com/LostRuins/koboldcpp
SillyTavern, and a lot more out there if you look around.
For image generation there's obviously Stable Diffusion. A few different self-hostable frontends and backends, IE:
https://github.com/AUTOMATIC1111/stable-diffusion-webui
https://github.com/comfyanonymous/ComfyUI
And https://civitai.com/ for a lot of your SD resources.
Obviously there's a LOT more AI tools out there than just this, I'm just listing off some of the core starting points.
Much obliged, thanks for typing all that out. Would be very interested in a 101 thread if you ever find the time, and I'm sure plenty of others here would be too.
Thanks, and I'm glad you gave me the nudge I needed to write some of that up. I'll definitely try to put something together for a more comprehensive loose 101 type of thread. I'm sure AlfredicEnglishRules and a few others could add a fair share of info as well.
And I'll preface that I by no means consider myself an expert on the subject. I was just lucky that I knew a person who'd been dabbling with it early on, and came across a few solid guides that helped get me started. Had to do a lot of extra digging from there though since so much stuff is so poorly documented.
I actually learned a lot from your other random comments in the past. Maybe just collect and organize some of those together in one post.
Thanks, I really appreciate that, and I'm glad I could help provide some insights.
I did take your suggestion and tried digging up some tips I've made in previous comments. Judging by a search in my comment history, I don't think I've generally gone into detail on any specific tools, guides, or basic setup info. I mean prior to a few comments in this thread.
I could probably whip something up pretty fast in the next day or so though. There's some stuff I'd like to expound on a bit more from my previous comment, some stuff I'd skipped or missed, etc.
You could check out the mrdeepfakes forums. There you will probably find some links to some github links and even some training material.
I will make a guide in a few days
Looking forward to this.
There's way too much noise and clickbait blogs in the search results for a lot of this AI stuff right now. It'll be nice to have some reliable reference material.
And a lot of that shit is probably AI generated itself.
That, and I suspect that large companies who are invested in this (especially those who own the search engines in question) are deliberately scattering search results to hamper competition.
But then again, we've also seen how companies like Google have almost completely stopped bothering to do any kind of work on managing their broken search engine. It's to the point to where I'll sometimes use AI to look up basic "common knowledge" information instead.
Probably using a fine tune of SD 1.5 with a Taylor Swift LoRA