• Member Since 2nd Mar, 2013
  • offline last seen 2 hours ago

SweetAI Belle


Why does life always have to be so ironic?

More Blog Posts80

  • 3 weeks
    The AI Pony Library Trip

    Got a new graphics card that generates everything faster (RTX 3060 12GB).

    Read More

    4 comments · 42 views
  • 10 weeks
    Mastodon/Twitter

    Just thought I'd note I'm slowly winding down my Twitter account. I haven't logged out of it yet, but I'm mostly just watching to make sure I get everyone's new account information for other places and watching the trainwreck. I'm still curious how badly things will go when something breaks and all the people that could fix it were fired or left...

    Read More

    0 comments · 42 views
  • 12 weeks
    New pony model on AI

    Since I've been blogging about image generating ai, thought it might be worth mentioning that a new version of the pony model came out.
    https://huggingface.co/AstraliteHeart/pony-diffusion-v2




    Read More

    3 comments · 34 views
  • 13 weeks
    Pony pony pony pony AI

    I said in my last blog post that I downloaded and installed Stable Diffusion, an AI image generator thingy. I'm obviously still playing with it.

    Read More

    7 comments · 77 views
  • 14 weeks
    Adventures in AI Image Generation

    I've been interested in trying out ai image generation for a while, ethical questions aside. While there are plenty of ai image generators online, they usually have limitations or want to charge you money eventually, and I thought it'd be more fun to set it up on my own computer, if it'd run.

    Read More

    4 comments · 73 views
Oct
28th
2022

Pony pony pony pony AI · 1:16am Oct 28th, 2022

I said in my last blog post that I downloaded and installed Stable Diffusion, an AI image generator thingy. I'm obviously still playing with it.

It runs off training models. While there's a main one you are supposed to download, there are a bunch of others, too. Big one I was playing with is a pony sfw model, trained on a lot of sfw pony pictures that had more than 500 upvotes or something. There are a bunch of others, though, including some furry ones that seemed to have ponies in them. Catch with those is they aren't sfw, so you have to specify characters wearing clothes and such.

Sooo... when I noticed you could merge models, I made one that's the pony database merged with the furry ones. :unsuresweetie:

But, while typing in prompts and making the characters do your bidding is fun, I was curious what it'd do if I wasn't giving it anything to work with. There's a "CFG Scale" setting from 1-30 that tells it how much attention to give the prompt, and I tried setting that to 1 and not typing in a prompt. Which turned out really boring and non pony related.

So, instead, I set that back to 7, and gave it a prompt of "pony pony pony pony". 'Cause pony was really what I wanted. Here's the first run, set to output 4.
#1:

Top left was kinda generic. Top right was cute, though one pony is missing a leg. Missing or extra body parts is always an issue, though. At least that pony didn't have a leg for a tail (which I saw at one point!). Bottom left has a great expression, though her mane can't decide whether it wants to be a mane or a hat. And I quite like the bottom right pony, though I do see an extra hoof. Looks kinda familiar for some reason. I always wonder if it's occasionally mostly turning up someone else's picture...

#2:
So I told it to give me 9 more.

Top left is kinda Discord-y, though rather muddled. Top middle looks kinda messed up. Think the text is nonsense. Most text in AI pictures is. It's one of its weak points. Top right is generic Applejack-y pony. Middle left actually looks pretty nice. Kinda Roseluck vibes from one of them. It messed up the tails, tho'. Middle is, um, memorable? Middle right isn't bad. Four ponies, two of which are Applejack-y types. Bottom left is nice aside from the tail issue. Bottom middle, I like, though you can see the fake text again. Bottom right is messed up, but I can see a good drawing in there wanting to get out.

#3:
I ran it one more time, with sampling steps cranked up, making it take three times longer. This often gives better results, though this time...

Top left... reminds me of nothing so much as paper jam Dipper in Gravity Falls. Top middle almost looks like someone repeatesly trying to draw the same character and messing up. Top right is kinda striking, actually. Middle left is missing a leg, but I quite like it. Middle is silly. Middle right is also silly with Applejack "Whooo! Whooo!" vibes. Bottom left is random pony sketch paper vibes. Bottom middle is actually pretty good, aside from missing and extra leg mixups, eye mishaps, etc. I rather like the look of the bottom right, though it has similar issues.

All in all, it was interesting to see what it comes up with. I sometimes feel like it'd be fun to redraw some of the better ones myself, or at least edit them to get rid of the issues.

Mostly, I had some fun looking over the results, and though others might too!

Comments ( 7 )

...Wow!

So many of these look... almost right, but then your eyes catch on something- usually any mouth that's not just a line- and suddenly, the whole pot's kind of spoiled. There are the ones that look too good- i'm thinking of the bottom left of the first batch- which makes me think it leaked in some kind of actual picture.

And then there are the messes. Like... mooost of them are kind of messes in little ways. Like melty ponies run over a dozen times with a blowtorch, and then they had their tails plucked out. :twilightsheepish:

This is still all very interesting to look at, though! After all, AI is both fun to mess with and to make fun of! This combines the best of both, whether you get total crap or immaculate construction!

...And in the end, maybe AI just isn't ready for pony?

Maybe AI will never be ready for pony...

I mean, who knows what would happen if AI got tangled up with ponies?

5694866
Yeah. Mind you, it doesn't help that I was using a merged model and was giving it very little to work with. It does know some pony names, and it'll sometimes follow a prompt fairly well, or miss individual bits. I think there are keywords you can add to your prompt to improve the output, too. But it does, also, sort of come down to picking out the good results and discarding the rest.

It really is "some" names, too:

Twilight Sparkle sitting at a table eating a sandwich.

i.imgur.com/d8UIWuY.png

Spike sitting at a table eating a sandwich.

i.imgur.com/U4qAsJT.png

It could also partially be the model. Here's wd 1.3, with a non-pony prompt (specializes in anime waifus, I think?):

A cute girl with brown eyes, white hair, and a blue dress floating in the air holding a glowing ball.

i.imgur.com/IChKrOt.png
(But then, that's a pretty cliche prompt that anime has tons of...)

There are other things you can do, too, that I haven't explored much. You can feed it both a prompt and a picture, and have it generate a new picture based on the old one following the prompt. And then you can keep doing it, using the best of the generated ones as the new picture to generate off of until you get what you want. Haven't really played with that yet, though.

One thing it does make me think is that using ai art instead of an artist isn't totally going to happen, because you need an artist to manage the ai to some degree. I think it'll be more of a tool an artist uses...

--Sweetie Belle

5694884
Haha-wow! I think I'll be seeing some of those Spikes in my dreams tonight! Hahahahahahahahaha....

mmmyeah, AI and technology in general are the future and such, but I think it'll take a very long time until something very essential about most beloved art is distilled into a formula. It's the minutia- the heart- and the soul. Little bits of beautiful detail like an oblique reference, or a scenic interpretation of something precious.

It's not even that the much more developed anime girl generator not bitter stomps on my point- like i thought it did when I started writing this reply- even those relatively beautiful and normal-looking pictures are lacking something; specifically: they are really beautiful and normal-looking. But that's all they are. Just, generic, formulaic anime art that's completely stunning to behold, but, as made evident, could theoretically be churned out ad nauseum by a machine. It's so obviously pretty-looking, but it all lacks soul. There are little details, but there's no minutia. There's nothing to recognize.

...Now, as a tool, then sure, maybe. If, theoretically, an artist- or in this case, a writer- were to input every single detail of a desired piece down to the finest detail as instructions, then an AI would probably be able to copy something relatively soulful.

...Because somebody told it about all the little, soulful tidbits to render. I'm not saying inanimate, unfeeling software doesn't have the capability to, maybe, possess the ability to create something definably 'human' as a contrivance- and whether that's at all distinguishable from what could be called contrived creations of the ill-defined 'mind' in the first place- but I just don't think we're there yet?

...I got a little too wordy for 11pm. ummmmmmm- technology's cool and i don't know anything about it i just like words have a good night sweetai belle.

I'm vaguely reminded of how I made a joke once or twice about a band called "Pony! Poni! Poné!" (instead of "Tony! Toni! Toné!").

It's not even that the much more developed anime girl generator not bitter stomps on my point- like i thought it did when I started writing this reply- even those relatively beautiful and normal-looking pictures are lacking something; specifically: they are really beautiful and normal-looking. But that's all they are. Just, generic, formulaic anime art that's completely stunning to behold, but, as made evident, could theoretically be churned out ad nauseum by a machine. It's so obviously pretty-looking, but it all lacks soul. There are little details, but there's no minutia. There's nothing to recognize.

I'd also keep in mind that if you examine those anime girls, you'll find issues, too. Look at the fingers, especially. One of the dresses just kinda fades out, too. It's just so pretty you don't notice the problems immediately.

But one thing is that it's all based on a large collection of images that that particular model is trained on, and logically, the more images it was trained on that are similar to your prompt, the less trouble it'll have. Which is pretty much going to lead to it being best used for stuff that's fairly unoriginal, which is not something I'd particularly want.

...Because somebody told it about all the little, soulful tidbits to render. I'm not saying inanimate, unfeeling software doesn't have the capability to, maybe, possess the ability to create something definably 'human' as a contrivance- and whether that's at all distinguishable from what could be called contrived creations of the ill-defined 'mind' in the first place- but I just don't think we're there yet?

We're probably not there yet for something that's actually original. If you are doing a game with generic anime protagonists, and you need 3-4 portrait pictures of them to put next to their dialog, you could probably manage it. Or you could take existing old low res artwork and get ai to rework it into higher resolution versions. It has uses, but I think if you just churn out ai artwork and put it in, say, a game blindly, you'll get something like the recent Grand Theft Auto remasters.

--Sweetie Belle

5694893
I'm pretty sure I had a song in mind when I wrote that prompt, but it's escaping me.

For what it's worth, my attempts to get the AI to do Twist went something like this:
i.imgur.com/w2dkrNq.png

It doesn't know her name, so I had to describe her, then it'd get details wrong...

--Sweetie Belle

Login or register to comment