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

«‹›435 »

Sedgr: Sep 16, 2007; Neat!

Sedgr fucked around with this message at 15:33 on Nov 7, 2022

# ? Oct 21, 2022 21:42

Adbot: ADBOT LOVES YOU

# ? May 30, 2024 10:59

mobby_6kl: Aug 9, 2009; by Fluffdaddy

I see your schwartz is almost as big as mine!

# ? Oct 21, 2022 21:44

Mr Luxury Yacht: Apr 16, 2012

Tried 1.5. It's hard to tell if it's better to be honest. Maybe slightly better at hands? Still getting a fair amount of fleshmonsters though.

# ? Oct 21, 2022 23:12

BoldFace: Feb 28, 2011

Lexica.art seems to have added free image generation to their site. I remember them receiving hefty amount of funding a while back. I wonder what their eventual business model will look like.

# ? Oct 21, 2022 23:29

Elotana: Dec 12, 2003; and i'm putting it all on the goddamn expense account

WhiteHowler posted:

Now that the SD source is out there, training a new model only takes time and processing power. This seems like an opportunity for another company to step in and offer an uncensored model.

You're going to need to wait for this tech to get substantially cheaper before it gets within reach of the kinds of companies willing to do that.

# ? Oct 22, 2022 00:36

WhiteHowler: Apr 3, 2001; I'M HUGE!

Elotana posted:

You're going to need to wait for this tech to get substantially cheaper before it gets within reach of the kinds of companies willing to do that.

I'm not so sure.

Look at StabilityAI -- in two months they went from virtual unknowns outside of the tech-nerd AI space to being the perceived leader in the next cutting-edge technology that will change the Internet. There are suddenly front-page articles about them on every tech site and interviews with the founder in the New York Times and on CNN. They've raised $101 million in capital and are valued at $1 billion. There are dozens of other start-ups out there hungry for that. Putting out a good, open model that eclipses SD could catapult someone else into that spotlight.

SD 1.4 allegedly took about $600k to train. That's pocket change when you're talking about tech companies chasing a hot new technology.

We'll see. This stuff is moving ridiculously fast right now.

# ? Oct 22, 2022 01:34

BrainDance: May 8, 2007; Disco all night long!

I wonder how expensive training will be in the long run? Will the needs of the technology scale with the improvements in hardware? Or will we just have much cheaper training a decade from now when the hardware requirements to do the training don't seem as absurd?

I am amazed no one has decided to be the people that explicitly make "the porn model." I'm sort of of the opinion that all this hand wringing about the dangers of AI, both in making fake images to pass them off as real, and of all the wild dirty porn they can make, are pointless because no matter what you think it's here and it's happening, cats out of the bag.

But I also kinda get why everyone is "fighting" it, because even though it's an inevitability, someones gonna eventually be the first person to get sued for it all and no one wants to be that. But drat, porn model trained with absolutely no names of anyone, do it and make millions? It's gonna happen, whether it happens because someone decides to profit off it right now or because training gets cheaper and anyone's able to do it.

# ? Oct 22, 2022 01:44

lunar detritus: May 6, 2009

BrainDance posted:

I am amazed no one has decided to be the people that explicitly make "the porn model."

I think it would have immediately been made the day after SD was released if there was any tagged database. No database, no training, and if training itself is relatively expensive for normal human beings, making a database is 100% out of reach.

# ? Oct 22, 2022 01:48

BrainDance: May 8, 2007; Disco all night long!

lunar detritus posted:

I think it would have immediately been made the day after SD was released if there was any tagged database. No database, no training, and if training itself is relatively expensive for normal human beings, making a database is 100% out of reach.

I had assumed porn would be the world's most tagged thing, but maybe that's just for videos? Or maybe it's not all centralized?

I also wonder about AIs tagging images themselves. All I know about that is apparently Facebook can do it (they tell you what's in a picture even if it wasn't tagged manually, like "2 people in a canoe laughing") which, if that can improve and becomes accessible, might just solve the whole problem.

BrainDance fucked around with this message at 01:56 on Oct 22, 2022

# ? Oct 22, 2022 01:52

WhiteHowler: Apr 3, 2001; I'M HUGE!

BrainDance posted:

I wonder how expensive training will be in the long run? Will the needs of the technology scale with the improvements in hardware? Or will we just have much cheaper training a decade from now when the hardware requirements to do the training don't seem as absurd?

It will get cheaper and faster, but the current limitation is the quantity and quality of tagged images to train on. The LAION 5b dataset is fully tagged, but if you go digging in there, many of the tags are vague, misleading, or outright incorrect.

lunar detritus posted:

I think it would have immediately been made the day after SD was released if there was any tagged database. No database, no training, and if training itself is relatively expensive for normal human beings, making a database is 100% out of reach.

If horny Tumblr were still around, I think this would have happened already. The images were generally well-tagged by the community, and the API was easy to query. Not sure if this is still the case, and they went (generally) SFW a few years ago anyway. My wife informed me that they just brought back boobs?

BrainDance posted:

I also wonder about AIs tagging images themselves. All I know about that is apparently Facebook can do it (they tell you what's in a picture even if it wasn't tagged manually, like "2 people in a canoe laughing") which, if that can improve and becomes accessible, might just solve the whole problem.

This is already possible and happening, but the problem with self-training is that there's no outside confirmation of correct/incorrect assumptions. When an AI looks at a photo of a 1954 Chrysler Crown Imperial and decides "this is a boat", if nobody tells it otherwise, its perception of what a boat looks like gets skewed.

# ? Oct 22, 2022 02:04

WhiteHowler: Apr 3, 2001; I'M HUGE!

Double-posting but who cares.

There's a neat tool called SD Infinity that offers a web-based UI specifically for outpainting on an infinite canvas. Give it a prompt, then drag the active canvas around and use the outpainting function to make a more complex or varied scene than you normally could via basic SD outpainting.

There's a demo here:
https://huggingface.co/spaces/lnyan/stablediffusion-infinity

It's slow because you have to wait in the SD demo queue, but the possibilities quickly become apparent. There's a locally installed version too, but I haven't tinkered with it yet.

Here's a thing I made using the web demo of SD Infinity:

It's kind of nonsense, but I like it. I could have made the overall image much more coherent with more outpainting retries, but I was tired of waiting on the SD demo site, so I generally took the first result.

# ? Oct 22, 2022 02:11

Methylethylaldehyde: Oct 23, 2004; BAKA BAKA

WhiteHowler posted:

It will get cheaper and faster, but the current limitation is the quantity and quality of tagged images to train on. The LAION 5b dataset is fully tagged, but if you go digging in there, many of the tags are vague, misleading, or outright incorrect.

If horny Tumblr were still around, I think this would have happened already. The images were generally well-tagged by the community, and the API was easy to query. Not sure if this is still the case, and they went (generally) SFW a few years ago anyway. My wife informed me that they just brought back boobs?

This is already possible and happening, but the problem with self-training is that there's no outside confirmation of correct/incorrect assumptions. When an AI looks at a photo of a 1954 Chrysler Crown Imperial and decides "this is a boat", if nobody tells it otherwise, its perception of what a boat looks like gets skewed.

Yeah, the various borou sites with the incredibly detailed tags for every single picture is what got the huge jump in output quality from the anime training set. Collecting and Tagging data correctly and accurately is the almost impossibly expensive part of most of these processes.

I do wonder how long it'll take before someone has a temporal filter than constrains changes between frames, so you don't have that pukefest A Scanner Darkly rotoscope effect going on. Mostly because I wanna heat my home this winter, and have a Studio Ghibi remake of Hellraiser or Aliens as my useful work product.

Edit: That or something that is able to convert Dwarf Fortress art descriptions and historical events into art. Use the history log as training/prompt data for a text generating AI, then feed the ingame art and item descriptions along with the context created by the text AI to make an illustrated manuscript of the insanity spiral that was Boatmurdered.

Methylethylaldehyde fucked around with this message at 02:17 on Oct 22, 2022

# ? Oct 22, 2022 02:14

RPATDO_LAMD: Mar 22, 2013; 🐘🪠🍆

BrainDance posted:

I wonder how expensive training will be in the long run? Will the needs of the technology scale with the improvements in hardware? Or will we just have much cheaper training a decade from now when the hardware requirements to do the training don't seem as absurd?

I am amazed no one has decided to be the people that explicitly make "the porn model." I'm sort of of the opinion that all this hand wringing about the dangers of AI, both in making fake images to pass them off as real, and of all the wild dirty porn they can make, are pointless because no matter what you think it's here and it's happening, cats out of the bag.

But I also kinda get why everyone is "fighting" it, because even though it's an inevitability, someones gonna eventually be the first person to get sued for it all and no one wants to be that. But drat, porn model trained with absolutely no names of anyone, do it and make millions? It's gonna happen, whether it happens because someone decides to profit off it right now or because training gets cheaper and anyone's able to do it.

NovelAI allows porn, and even has a nsfw toggle on their web interface. They don't market it though. People who have seen the file list from the leak will know that they have a separate sfw and nsfw model, trained differently
In terms of whole-identity-focused stuff the people who own the Unstable-Diffusion discord (who partially funded the finetuning for the waifu diffusion 1.3 anime model) are currently running a patreon to develop their own porn-focused AI website.
They've posted WIP designs for the frontend:

BrainDance posted:

I also wonder about AIs tagging images themselves. All I know about that is apparently Facebook can do it (they tell you what's in a picture even if it wasn't tagged manually, like "2 people in a canoe laughing") which, if that can improve and becomes accessible, might just solve the whole problem.

Already happened.
Stable diffusion was trained on the LAION-5B data set which consists of 5 billion images tagged by the CLIP AI. If you use the AUTOMATIC webui there is even an "Interrogate CLIP" button to get a CLIP description of any image. For example for your avatar it produces "a turtle with a sun in the background and a tree in the foreground with a blue sky and clouds"

RPATDO_LAMD fucked around with this message at 02:27 on Oct 22, 2022

# ? Oct 22, 2022 02:18

SubG: Aug 19, 2004; It's a hard world for little things.

mobby_6kl posted:

The Novel model is generally pretty good about hands in normal images but if you try to make it do closeups, it's all absolute nightmare sausage fingers

I don't know about the anime models, but for vanilla 1.4 one of the best ways I've discovered for getting the details right, hands or otherwise, is to add poo poo near the end of the prompt like you're a professional photographer adding tags to a photo in your flickr. So the prompt ends up being something like "[subject of image], shot with [random specific model of DSLR], [lens spec], natural lighting". If you want something that looks like it was shot on high ISO black and white film stock (so kinda grainy but still a little soft) something like "[subject of image], shot with Nikon D780, 85mm, chiaroscuro, dramatic lighting, cinematic" seems to work. Even when it swings and misses on hand anatomy, it usually looks "right": like a photorealistic image down to individual pores on the skin and so on, only with two pinkies or something. That's usually pretty easy to clear up via inpainting.

# ? Oct 22, 2022 02:24

frumpykvetchbot: Feb 20, 2004; PROGRESSIVE SCAN; Upset Trowel

WhiteHowler posted:

the current limitation is the quantity and quality of tagged images to train on. The LAION 5b dataset is fully tagged, but if you go digging in there, many of the tags are vague, misleading, or outright incorrect.

This is already possible and happening, but the problem with self-training is that there's no outside confirmation of correct/incorrect assumptions. When an AI looks at a photo of a 1954 Chrysler Crown Imperial and decides "this is a boat", if nobody tells it otherwise, its perception of what a boat looks like gets skewed.

LAION 5b is huge so it is a daunting proposition to apply qualified human correction on every image-tagset pair. If someone threw a hundred million dollars at menial "Mechanical Turk" tasking to get this done, it would probably still not be enough to make a difference. I think investing more in improving self training is important. Maybe the self-tagging can at least come with some "doubt" or "fidelity" signal that would perhaps reduce the amount of distortion self-tagged images do to the dataset. If the self-tagging can't reliably discriminate a boat from a car, could it perhaps spit out weighted tags: Boat 80% car 15% art deco radio 5%

The pre-SD release historical image dataset must be preserved, archived. Post-SD releaase internet as tagged image source is now very unreliable, since so many people are sharing and posting AI generated images everywhere with prompt word salad attached that could be confused for tags suggesting that Greg Rutkowski singlehandedly painted tens of milions of images uploaded.

# ? Oct 22, 2022 02:34

Tunicate: May 15, 2012

frumpykvetchbot posted:

LAION 5b is huge so it is a daunting proposition to apply qualified human correction on every image-tagset pair. If someone threw a hundred million dollars at menial "Mechanical Turk" tasking to get this done, it would probably still not be enough to make a difference.

also google tried that on text and got a 30% failure rate

https://www.surgehq.ai/blog/30-percent-of-googles-reddit-emotions-dataset-is-mislabeled

mostly because they mechanical turked it and didn't bother getting people with a good grasp of idiomatic english

Tunicate fucked around with this message at 03:53 on Oct 22, 2022

# ? Oct 22, 2022 02:36

BrainDance: May 8, 2007; Disco all night long!

Well there you go, then that's the whole thing.

The current limitation is lack of tagged images, but AI itself can and does tag images. The AI can make mistakes tagging images, but humans also make mistakes and are apparently pretty bad at it given the google thing. Perfect is the enemy of good, if the AI mistakes 1 in a 1000 Chryslers as boats that's probably still better than people and a lot cheaper.

So, then that doesn't sound like there is a problem. Other than making the AI better at tagging images itself, we just do that... a lot. I guess instead of training the AI on all the porn out there we just get an AI to tag all the porn out there first and make the specifically porn AI that only does porn and get rich.

Or other non-porn things, I just think the porn is the obvious direction for this to all go. I think a lot of the utility of AI image generation can come from single purpose models trained on very specific things and only those things, like I mentioned before I think, no one would do this and there are problems doing it that other sets of training data wouldnt have but I think it would be cool, a sprite art model trained on all the NES/SNES/Genesis etc sprites.

# ? Oct 22, 2022 03:28

LifeSunDeath: Jan 4, 2007; still gay rights and smoke weed every day

you guys sure are talking about porn a lot lol.

# ? Oct 22, 2022 03:58

BrainDance: May 8, 2007; Disco all night long!

LifeSunDeath posted:

you guys sure are talking about porn a lot lol.

Really just me, and I don't actually even care about AI porn itself just, that's where a lot of the criticism of and attention on AI image generation is at.

And where I think the biggest, and inevitable, problems in the future will be. Deepfakes exist and are already a problem, and it's all just going to get better at making porn regardless of what anyone does.

# ? Oct 22, 2022 04:10

frumpykvetchbot: Feb 20, 2004; PROGRESSIVE SCAN; Upset Trowel

Tunicate posted:

also google tried that on text and got a 30% failure rate

https://www.surgehq.ai/blog/30-percent-of-googles-reddit-emotions-dataset-is-mislabeled

mostly because they mechanical turked it and didn't bother getting people with a good grasp of idiomatic english

if the UK is left to implode under another decade of tory govt, it will be a third world country with millions of perfectly qualified English speakers sufficiently destitute to take on such tasks for cryptotokens that can be bartered for food rations.

# ? Oct 22, 2022 04:24

Mercury_Storm: Jun 12, 2003; *chomp chomp chomp*

"a gigantic killer kaiju monster ball of yarn with giant cute googly eyes and large fangs attacking the city of Portland Oregon and destroying buildings"

"the cover of a storybook book titled "doot doot doot doot doot doot doot doot" about a baby harp seal that plays the trumpet"

Mercury_Storm fucked around with this message at 05:18 on Oct 22, 2022

# ? Oct 22, 2022 04:42

LifeSunDeath: Jan 4, 2007; still gay rights and smoke weed every day

BrainDance posted:

Really just me, and I don't actually even care about AI porn itself just, that's where a lot of the criticism of and attention on AI image generation is at.

And where I think the biggest, and inevitable, problems in the future will be. Deepfakes exist and are already a problem, and it's all just going to get better at making porn regardless of what anyone does.

I'm just messin, no worry yo. AI porn is gonna be wild af, I have no clue what that's gonna end up like.

# ? Oct 22, 2022 05:07

Squatch Ambassador: Nov 12, 2008; What? Never seen a shaved Squatch before?

https://www.youtube.com/watch?v=8EE54ar7CrQ
Make sure to set it to 1080p, YouTube's bitrate at lower resolutions can't handle this nightmare.

The first several shots used the prompt "masterpiece, best quality, Goku SSJ2 is spying on Kakarot Arctor, HD", with seed, settings, and hypernetworks changed between shots. At some point I got bored and started adding other DBZ characters and Anime. Mostly done using Novel AI, and one shot with the Studio Ghibli model. The Security camera footage used the prompt "Goku at home, security camera footage, dragon ball z". I couldn't get it the extreme closeup of Keanu's face to look good, so I cranked up the denoising strength and let it make random Gokus.

# ? Oct 22, 2022 05:26

Pac and Cheese: Oct 29, 2010; gotta walk fast

tried making this and dall-e mini banged it out perfect the first try

# ? Oct 22, 2022 09:59

Megazver: Jan 13, 2006

LifeSunDeath posted:

I'm just messin, no worry yo. AI porn is gonna be wild af, I have no clue what that's gonna end up like.

realistically, it's going to be all the celebs, doing all the things you ever wanted them to do in front of you, you horny horny doge

also, sigh, a whole lot of WEIRD porn

# ? Oct 22, 2022 10:39

Rutibex: Sep 9, 2001; by Fluffdaddy

I can't wait to make Mr bean and He-man crossovers

# ? Oct 22, 2022 10:48

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

Rutibex posted:

I can't wait to make Mr bean and He-man crossovers

Why are you putting Mr Bean's face on Heman's body?

# ? Oct 22, 2022 10:54

Boba Pearl: Dec 27, 2019; by Athanatos

Why do people think Disney is this magical entity? Like it's not a fae trickster, where if you say it's name 3x a lawyer comes out and puts two in the back of your head. People are saying "Hahaha, we'll defeat those dastardly AI Bros by training models on DISNEY and DISNEY will do it," but like... A: Distributed Networking exists, and has been proven to be successful on training AI, so they'd have to chase down... Thousands? Of people all over the internet, B: Disney can't stop fanart, porn, or pirating, so why is this the software it'll magically have power over, C: The Rule34 model was like one of the first ones to come out and D: Why is it always disney?

# ? Oct 22, 2022 10:56

pixaal: Jan 8, 2004; All ice cream is now for all beings, no matter how many legs.

Once you can add training data using consumer hardware that's under say $300 this will really take off.

What if the real trick isn't to tag existing art but make art specifically that will tag well that is a solid training set? This would likely be a generation's work but is it what is needed to remove the noise.

# ? Oct 22, 2022 11:01

Tamba: Apr 5, 2010

pixaal posted:

Once you can add training data using consumer hardware that's under say $300 this will really take off.

What if the real trick isn't to tag existing art but make art specifically that will tag well that is a solid training set? This would likely be a generation's work but is it what is needed to remove the noise.

we could even pay the (now starving) artists to do so :v:

# ? Oct 22, 2022 13:01

Megazver: Jan 13, 2006

I suspect we'll soon see an effort to start a crowdsourced Human Art Tag Project, where thousands of horny autists are going through Pinterest and tagging every single pic, like, "big tiddy goth gf, horny, 2d, not greg rutkowsky, soft lighting, normal human anatomy"

# ? Oct 22, 2022 13:34

precision: May 7, 2006; by VideoGames

Megazver posted:

I suspect we'll soon see an effort to start a crowdsourced Human Art Tag Project, where thousands of horny autists are going through Pinterest and tagging every single pic, like, "big tiddy goth gf, horny, 2d, not greg rutkowsky, soft lighting, normal human anatomy"

a website that employs artists 24/7 to manually fill DALLE style requests

rooms full of artists chained to desks, sweating

# ? Oct 22, 2022 13:54

mobby_6kl: Aug 9, 2009; by Fluffdaddy

precision posted:

a website that employs artists 24/7 to manually fill DALLE style requests

rooms full of artists chained to desks, sweating

# ? Oct 22, 2022 14:02

Vlaphor: Dec 18, 2005; Lipstick Apathy

Vlaphor posted:

I finally got that new script working (didn't realize it was a gui and that you just ran the script), but yeah, I'm digging this new system.

"The gateway between dreams and nightmares, sinister magical vibe, 4k, realistic, vray lighting"

Now that I'm better at this, I thought I'd retry one of my previous prompts with some new settings and 1.5

"The gateway between dreams and nightmares, sinister magical vibe, 4k, realistic, vray lighting"

I love how it does gateways and portals

# ? Oct 22, 2022 14:02

precision: May 7, 2006; by VideoGames

"beavis and butt-head, realistic photograph"

# ? Oct 22, 2022 14:10

mobby_6kl: Aug 9, 2009; by Fluffdaddy

precision posted:

"beavis and butt-head, realistic photograph"

SD seems to have absolutely no idea who Beavis and Butt-head are but the results are... something

"Holy Diver" on the other hand is pretty cool. It knows it's a name of an album.

# ? Oct 22, 2022 14:24

Mescal: Jul 23, 2005

TIP posted:

I just used Stable Diffusion for a professional project for the first time and it's pretty sweet.

I put together this capsule image for my game's Steam page and it's mostly AI art:

I started by posing some of my assets in game and taking a screenshot:

I put that into img2img and... it was a real mess. Way too much going on, so I rendered out each component separately on plain backgrounds and fed those into SD and then composited them back together. The trickiest part was the clown, I had to do his lower body separate from the top and generate the arms separately from the torso.

The environments are a combo of AI generated stuff closest to the camera and actual game render for the farther off details. I also did my own rim lighting on the hand and clown to better separate the elements.

I'm pretty happy with it, will probably use SD for most complicated art projects from now on.

Oh, and since the steam capsule is kinda small but I made all the elements huge, here's the clown and hand at high res: