AI Generative Artwork

The Something Awful Forums > The Finer Arts > Creative Convention > AI Generative Artwork

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

Welcome to the AI-art thread.

AI art, and AI in general, is literally going to change the world. The way it works is a "prompt" is written, and the AI model interprets the "prompt," and creates an image.

This technology is going to put power into the hands of idiots like me, and also into the hands of idiots in executive board rooms. It's going to change the way we work, the way we create, the way we go to school, everything. It's highly exploitable. It's also super fun. If you're curious about the tools, then this is the thread for you. Maybe I'll flesh out this post with a bit more detail, maybe I won't, who knows? But here's some neato information for you to begin the exploitation of this fun little field.

Web-Based:

Craiyon (free)

This is probably the most popular AI image generator out there. It's free, it's easy to use, and it creates whacky results.

This is "a cat dressed up like Chairman Mao giving a speech at the Pentagon." (mine)

Dall-E 2 (Free to try once you're in the beta, then not free)

Dall-E 2 is proprietary bullshit that costs actual real dollars in the form of credits, but it does produce neat results. It's based on OpenAI, which isn't "open" at all. If you've seen an article on the internet about the magic of AI generated images, it was probably focused on Dall-E 2.

This is "A photorealistic render of a cute fuzzy orange monster sitting at a desk in a depressing cubicle doing boring accounting work, path tracing, volumetric lighting, moody." (mine)

Discord-Based:

Midjourney (Free to try, then not free)

Midjourney is open to the public. It's free at first, but once you run out of free, it costs money. It's closed-source with no plans to make it open-sourced. It will soon have a web-based version. It's probably the biggest and most popular of the AI-generated image makers, even bigger than Dall-E in some ways. The results that it produces are clearly weighted to not produce "bad" results. People like using Midjourney because it requires less prompt engineering knowledge.

This is "Industrialpunk Fembot Natalie Kardashian with gears in her head" character portrait art by Ivan Azaizovsky, cgsociety" (not mine). You can easily find more examples of Midjourney on Google, or just go join their discord.

Stable Diffusion (Free, ~~and soon to be open source)~~
Stable Diffusion is the new kid on the block. ~~Their plan is to~~ They are now fully release and their models and weights are online, for free, forever. This is pretty cool. It produces results that are sort of a combination between Midjourney and Dall-E.

Currently, it exists only on Discord... unless you're very resourceful, then you can get a copy of it that'll run locally on your beefy GPU. It will soon officially be released to run locally. Currently, the local copy is only available to academics, and to people who know how to use google to find things.

Currently, you can run Stable Diffusion locally on your own GPU, or you can run it on their website. The website functionality is currently in a closed beta, but I think it'll open up soon. It does cost credits/money... but you can run it locally, for free, if you prefer.

All of these examples are mine:

"Amazing photorealistic digital concept art of a guardian robot in a futuristic city, by James Clyne and Joseph Cross. Cinematic. LED lighting. A bright billowing explosion in the distance. Wide angle. Clean lines. Balanced composition."

Local (note: Anything local requires a GPU with CUDA cores (sorry AMD), and at least 8gb of VRAM. 12gb is better. 24gb is more ideal)

Disco Diffusion

This is my personal favorite. The results are always a little surprising, it takes some finesse to get coherency, and it's never boring. It's also the slowest and the biggest pain in the rear end to get working.

Google Colab Version

A GitHub repository/Python/Anaconda/you're a developer version

Or a version that can run on a platform called Visions of Chaos, and the instructions on how to make it work are here. Visions of Chaos is an extremely powerful suite of software. It's free. It's amazing. The developer is super cool. The VoC Discord is awesome. If you're going to run Disco locally, this is the way to do it.

All of these examples are mine:

code:

"a beautiful hyperrealistic ultradetailed 3D render of a gigantic mecha guarding a dragon, by brian sum and stephen martiniere and Antonio Manzanedo:3",
            "mech, dragon, unreal engine, octane render, PBR, 3D, brilliantly colored, intricate, wide angle, volumetric lighting, polished, path tracing:2",
            "people, sphere, balloon, blurry, blur, depth of field, dof, bokeh, low quality, obscured, macro, jpeg artifacts, noise:-1"

Stable Diffusion ~~(SOON, or NOW if you're resourceful)~~
Stable diffusion is a free open source competitor to both Midjourney and DALL-2/OpenAI. It can be run locally, or it can be run via a web interface that, as of this writing, is still in limited beta access.

It can produce images quickly and with a high degree of fidelity.

Visions of Chaos can also be used to run Stable Diffusion, which is how I've been doing it and it's freaking great. This is the guide to getting the machine learning stuff up and running. Be prepared to fill up around 300gb of storage. I recommend putting it on a fast drive.

Stable Diffusion-UI is another front-end that allows you to easily install and run Stable Diffusion locally on your computer. It has a localhost-based WebUI. Looks pretty cool, though I haven't personally tried it yet.

Sneeze Party fucked around with this message at 15:30 on Sep 12, 2022

# ¿ Aug 18, 2022 15:08

Adbot: ADBOT LOVES YOU

# ¿ May 14, 2024 00:03

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

This space reserved, probably pointlessly.

Or not so pointlessly.

Here are some resources.

Lexica is basically a search engine for Stable Diffusion prompts. You want to make something trippy but you're short on time and/or you don't care about being original? Check this out.

CLIP Artists is a gigantic list of 600+ artists. It's a good place to start for artist style modifiers. "A painting of a cat in the style of Andreas Rocha" for example.

CLIP Prompt Engineering is a neat guide on the way that various styles and modifiers interact with one another. Very succinct information with visual examples.

A Traveler's Guide to the Latent Space is a Disco Diffusion-oriented guide. It's a fantastic read written by a guy who knows a lot about Disco. He's been using it since before most of us knew it existed. It explains every variable.

Clip Front is a search engine, of sorts, that will query the Laion5B dataset for whatever you want. For example, not sure if the AI knows who your favorite artist is? Plug their name into this search engine and find out for sure. Sometimes you have to put "by Artistname" for it to recognize what you're talking about. It's a neat tool.

Sneeze Party fucked around with this message at 14:56 on Sep 11, 2022

# ¿ Aug 18, 2022 15:08

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

You're absolutely right. It's going to get easier and easier to make things look better and better, and more how you want them to look.

This is one from last night, with Disco Diffusion.

Sneeze Party fucked around with this message at 14:21 on Aug 19, 2022

# ¿ Aug 19, 2022 14:04

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

JeffT417 posted:

https://stability.ai/blog/stable-diffusion-public-release

Here we go.

Updated the OP to reflect Stable Diffusion's release. I'm not a huge fan of it so far. I mean, it's neat, and it's fast, and you can get some clean looking results. But most of what I'm seeing out of it looks sort of generic.

# ¿ Aug 23, 2022 16:14

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

You can also run Stable Diffusion from within Visions of Chaos, which has been working really well for me. This is from a run I did last night. It's so fast on my GPU -- like a minute per render, or less.

Along with the rest of them:

# ¿ Aug 25, 2022 17:31

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

quote:

I�m really interested in what Disco Diffusion can do, but sadly haven�t had the time to get over the learning curve nor have had the patience to wait for its iterations - especially when MJ often takes literal seconds to output.

Disco Diffusion does take a long time, like 10 or 15 minutes per render for most people, but the results tend to be a lot more unique and surprising than the other platforms.

If you want to check it out, the easiest way is through Visions of Chaos. You can also use Stable Diffusion from within Visions of Chaos. SD is kind of like Midjourney in that it's fast and the outputs are really neat. Be aware that the total install size for Visions of Chaos is like 200 or 300 gigs. You're getting *a lot of stuff* to mess around with in it.

Midjourney does some fuckery on the backend to make nearly every output look mind blowing, but the result is that Midjourney output tends to have a very Midjourney look-and-feel. Regardless, it's a totally mind-blowing technology. In Stable Diffusion and Disco Diffusion, you maintain more direct control over your output.

Also, I'm updating the second post with some useful websites and resources.

# ¿ Sep 6, 2022 14:14

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

madmatt112 posted:

I gotta get more solid state storage

e: and a bigger GPU apparently lol

Stable Diffusion has forks now, which are available in Visions of Chaos, that'll run on 4-6gb of RAM. So any 8gb card will handily run it, and run it pretty darn well.

That being said, always get a bigger GPU

# ¿ Sep 8, 2022 14:30

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

Sitting Here posted:

Assuming you mean a wordpress blog, of course! I don't see a reason to not link it unless it contains something hateful or unsavory. And since I'm here--do you want a thread tag other than shitpost?

Great thread!

Thanks! Yeah, that'd be great. Don't know how I made the mistake in the first place. How about the ART tag?

# ¿ Sep 10, 2022 14:26

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

Those are seriously awesome. I've seen some other cool "schematic" styled generations, too. How are you running SD? Just through conda/or whatever, or Visions of Chaos?

Also, I updated the second post with some neat resources for anyone who is interested.

# ¿ Sep 11, 2022 14:57

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

ickna posted:

I'm using stable-diffusion-ui - https://github.com/cmdr2/stable-diffusion-ui#installation. It's a nice self-contained package that downloads everything automagically and then runs a web ui on localhost to interact with it, and is pretty quick to spin up the generator.

Cool, added this to the OP.

# ¿ Sep 12, 2022 15:30

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

Duuk posted:

I had no idea AI image generation had come this far. Amazing stuff.
Those Disco Diffusion naval/space action scenes are extremely my poo poo.

Thanks.

quote:

How does this work in terms of copyright? I got the impression from the Stable Diffusion site that it's "free cor non-commercial and commercial use," meaning you own the image you generate?

This is what is known as the grey area. The US Copyright Office has made statements indicating that AI-generated artwork cannot be copyrighted, but there are exceptions, and IANAL. Stable Diffusion, or rather stability.ai, wants you to agree to certain terms and conditions to use its services and its models. But I don't know how that works as being legally bound, goes.

quote:

If someone else uses the exact same keywords, will they get the exact same image, or is there some sort of infinite seed/iteration mathemagic going on in the background?

With Disco Diffusion, it is functionally impossible to recreate an image.

With Stable Diffusion, you just need to have the same resolution/prompt/model/seed.

quote:

How much time would it take studying this stuff to make something like the abovementioned spacey scenes if there's a particular scenario in mind? It'd make a hell of a book cover.

In Disco Diffusion, it is a much more labor intensive process to learn the hows of prompt engineering. The spelljammer/space pirate stuff took me no fewer than 50 separate prompts. And with Disco, that was like 10 or 15 minutes per render.

Stable Diffusion makes it a lot easier to get *good* stuff, but it still takes some knowledge to get *wow* stuff. Midjourney makes almost everything look somewhat dope, but it can't be run locally and it's not open source so it can go right on and gently caress off.

No offense to the Midjourney poster above. That's good stuff.

Sneeze Party fucked around with this message at 21:21 on Sep 12, 2022

# ¿ Sep 12, 2022 21:19

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

Doctor Zero posted:

Oh thank gently caress! Thank you for creating this, Sneeze Party. I wanted an AI thread, but I didn't want to have to create it.

I've been messing with Midjourney for a couple of weeks and it's pretty drat close to what I've wanted for ... forever, really. Good imagination + learning art = AI ART MOTHAFUCKA!

Glad to have other goons with me falling down the rabbit hole - like a lot of you my wife just tunes out when I start blathering about the 500 versions of a D&D character I made. I am getting in this heavily - I have a ton of ideas for uses for this (which I'll be doing well if I even use ONE).

I didn't know you can run this stuff local. I have a super beefy MacBook Pro 16" M2 - would it be better to build a dedicated windows () machine for this?

Some information on getting SD up and running on an M1 Mac. Not sure how helpful it'll be but feel free to post your results.

https://news.ycombinator.com/item?id=32678664

# ¿ Sep 14, 2022 01:39

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

Did you ever get it to work on your 1650?

# ¿ Oct 2, 2022 14:03

Adbot: ADBOT LOVES YOU

# ¿ May 14, 2024 00:03

Sneeze Party: Apr 26, 2002; These are, by far, the most brilliant photographs that I have ever seen, and you are a GOD AMONG MEN.; Toilet Rascal

If you're running Stable Diffusion locally, or at least are interested in doing so, I found this cool locally-run Discord bot. I've got it running on my private server for my friends. The instructions for installing it are pretty straightforward.

https://github.com/AmericanPresidentJimmyCarter/yasd-discord-bot

The whole thing in FOSS, which is nice. And the developer is super friendly, chill, and helpful. Attached is an example of what the bot looks like in a Discord channel.

Only registered members can see post attachments!

# ¿ Nov 14, 2022 19:59

The Something Awful Forums > The Finer Arts > Creative Convention > AI Generative Artwork