Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

[Pocket Dragoon]

Spoiler: Unreleased King's Quest, 1990

 

[crowabunga]

I've been having a lot of fun with Automatic1111's SD interface. I've been running it with the --listen argument on a PC with an RTX 3060, and then prompting from a laptop before going to bed. It's a comfier nighttime activity than doomscrolling, and cheaper than NovelAI (well, considering that I bought the RTX 3060 for other reasons).

 

[urr13 account]

They're mad the AI art is often better than the garbage they churn out and people who get commissions from them for porn can now get that shit for free.

Another POV from /pol/ I hadn't thought about:
View attachment 4150065

Also it's been memed about earlier, but I love to bring this point up repeatedly:
View attachment 4150071

Since this ai art boom I've seen so many incredible imaginative creations, could it be that the most creative people only now, thanks to AI, have the ability to bring their ideas to life.

All the truly creative and imaginative people have been trapped in day jobs, weighed down by responsiblites while the spoiled entitled brats who had all the free time in the world to learn drawing and pretend to be "artists" on the internet clearly lack any creativity or original thought.

All this talk of ai infringing on the human spirit is nonsense, really the opposite is happening the truly creative manifestations of the human spirit are now free'd thanks to ai cutting out the need for creative people to hire spoiled art school faggots who would probably fuck up the assignment anyway.

It's going to be really telling when we have ai that can generate feature films and we realize that some random guy in Nebraska, who would have never been able to get Hollywood funding or resourses to make a movie is able to create better cinema than 90% of Hollywood making them all look like baffoonish hacks with their embarassing unimaginative trash.

AI is freeing the real human spirit, pure creation, from those who weilded it through the happenstance of being able to live lives of leisure where they drew pictures all day or had some studio cut them a check to make garbage movies. This is only the begining.

 

[Balr0g]

Tried my hand on a pixel art castle

 

[crowabunga]

Since this ai art boom I've seen so many incredible imaginative creations, could it be that the most creative people only now, thanks to AI, have the ability to bring their ideas to life.

All the truly creative and imaginative people have been trapped in day jobs, weighed down by responsiblites while the spoiled entitled brats who had all the free time in the world to learn drawing and pretend to be "artists" on the internet clearly lack any creativity or original thought.

All this talk of ai infringing on the human spirit is nonsense, really the opposite is happening the truly creative manifestations of the human spirit are now free'd thanks to ai cutting out the need for creative people to hire spoiled art school faggots who would probably fuck up the assignment anyway.

It's going to be really telling when we have ai that can generate feature films and we realize that some random guy in Nebraska, who would have never been able to get Hollywood funding or resourses to make a movie is able to create better cinema than 90% of Hollywood making them all look like baffoonish hacks with their embarassing unimaginative trash.

AI is freeing the real human spirit, pure creation, from those who weilded it through the happenstance of being able to live lives of leisure where they drew pictures all day or had some studio cut them a check to make garbage movies. This is only the begining.

Reminds me of a part of "Captain Stormfield's Visit to Heaven":

“Did you ever see Napoleon, Sandy?”


“Often—sometimes in the Corsican range, sometimes in the French. He always hunts up a conspicuous place, and goes frowning around with his arms folded and his field-glass under his arm, looking as grand, gloomy and peculiar as his reputation calls for, and very much bothered because he don’t stand as high, here, for a soldier, as he expected to.”


“Why, who stands higher?”


“Oh, a lot of people we never heard of before—the shoemaker and horse-doctor and knife-grinder kind, you know—clodhoppers from goodness knows where that never handled a sword or fired a shot in their lives—but the soldiership was in them, though they never had a chance to show it. But here they take their right place, and Cæsar and Napoleon and Alexander have to take a back seat. The greatest military genius our world ever produced was a brick-layer from somewhere back of Boston—died during the Revolution—by the name of Absalom Jones. Wherever he goes, crowds flock to see him. You see, everybody knows that if he had had a chance he would have shown the world some generalship that would have made all generalship before look like child’s play and ’prentice work. But he never got a chance; he tried heaps of times to enlist as a private, but he had lost both thumbs and a couple of front teeth, and the recruiting sergeant wouldn’t pass him. However, as I say, everybody knows, now, what he would have been,—and so they flock by the million to get a glimpse of him whenever they hear he is going to be anywhere. Cæsar, and Hannibal, and Alexander, and Napoleon are all on his staff, and ever so many more great generals; but the public hardly care to look at them when he is around. Boom! There goes another salute. The barkeeper’s off quarantine now.”

 

[Stephanie Bustcakes]

Does Stable Diffusion 2.1 just always suck?

Pretty much. Apparently is works better with negative prompts. The dataset used is worse since they ditched the previous one because it wasn't open source, iirc. I'd expect they'll add more to it over time and future versions will get better and better. For now though I'm sticking to version 1.5.

 

[Pendulous Dong]

Made a movie poster.

 

[Sinner's Sandwich]

View attachment 4158966
Made a movie poster.

This looks like a new Castlevania game...I wish it was.

 

[Stephanie Bustcakes]

Guys, I love seeing the big galleries of images you make, but if your AI outputs images as 24 bit .png files, please resave them as .jpgs with reasonable compression before you upload them.

Stable diffusion generated this image as a 24 bit .png file. It was around 800 kb. re-saving it as a .jpg at 40% quality brought it down to 89 kb, almost one ninth the size, and if you looked at the two versions side by side you wouldn't notice any difference. I tried various compression levels and it was only in the 30-40% range that I noticed a drop in image quality, and even then just barely.

If you save all your images as .jpg files at 50% quality your posts will load 8x faster, which will let people see your images on slow connections and during ddos attacks. Plus it takes up less room on Null's server that he has graciously provided for us. (This should go for uploading 24 bit .pngs in general, not just for this thread)

Note: sometimes 8 bit pngs will be smaller and look better for things like line art, images with flat color like cartoons and MS Paint drawings, and monochrome images. See which works best for you, but 24 bit .pngs are always way too big.

 

[AmpleApricots]

Is midjourney still only reachable via discord bot? I want to spend some time on imagegen again and not sure if I should go the paperspace route again or try it. (yes, I'm aware of nai but it sounds like it's mostly for generating porn)

 

[ducktales4gameboy]

MJ is still discord only although they've been promising a web ui for ages.

NAI is fine for general Illustrations and nonphotorealism if you negative prompt for anime, but I'm going from experience only with their leaked checkpoint. The current version may be pornier.

 

[Monkey Pink]

I posted this to the AI Art Seethe thread:

https://kiwifarms.st/threads/ai-art-seething-general.128786/post-14032608

I don't want to double post but basically everything I said could go in this thread too.

Rundown:
-Artists still spewing the "AI is theft" lie but realize their argument is running out of steam
-Their current M.O. is to falsely accuse the Unstable Diffusion developers of being pedophiles
-4chan has had some interesting discussions in their "AI Art Ethics and Artist Seethe Thread" generals, archives are linked in above post.

Also, @Stephanie Bustcakes - your HD wojaks and other memes have gone super viral on /g/, congrats you're e-famous.

 

[WhiskeyJack]

Here's a question for those of you who can actually get this stuff to work; is it possible to use AI art generators to make something like a 3D render in Blender or Daz look more realistic? Because that seems far more useful to me than randomly inserting word prompts in hopes of getting something in the pose I want.

Cause right now if I type in "Black Widow riding a T-Rex' it spits out a dumb looking lizard wearing a black catsuit with wheels for feet. Not the most useful thing in the world.

 

[AmpleApricots]

Cause right now if I type in "Black Widow riding a T-Rex' it spits out a dumb looking lizard wearing a black catsuit with wheels for feet. Not the most useful thing in the world.

do some research into img2img, it's what you are looking for.

 

[Monkey Pink]

do some research into img2img, it's what you are looking for.

Is there any way I can do img2img for free on a shitty laptop or mobile? That's what I'm stuck with but I'd love to mess with it.

 

[Catgirls are Love]

Here's a question for those of you who can actually get this stuff to work; is it possible to use AI art generators to make something like a 3D render in Blender or Daz look more realistic? Because that seems far more useful to me than randomly inserting word prompts in hopes of getting something in the pose I want.

Cause right now if I type in "Black Widow riding a T-Rex' it spits out a dumb looking lizard wearing a black catsuit with wheels for feet. Not the most useful thing in the world.

Like stated above, Img2Img is where you're going to want to focus your efforts. I can guarantee that there were almost certainly zero images of people riding T-Rexs used in the training data of whatever model you're using, so good luck getting a person in a reasonable place on the dinosaur using text prompts. If you have a rough image to work from already that's a better starting point.

Also, the "riding" prompt may get horses/bicycles/vehicles mixed up. I haven't really tried to generate people riding horses (people on fantasy animals were attempted, didn't really turn out well - often just put the fantasy animal in frame but the actual person on a horse), but people on bicycles/tricycles occasionally have a horsehead show up just because.

Spoiler: a few early tests involving bicycles starring darth vader and palpatine

And these were the ones that turned out better; the model struggled enough with people on bicycles/tricycles and I assume it had plenty of training data to pull from. Didn't do any img2img trials that might've gotten better results.

 

[AmpleApricots]

Is there any way I can do img2img for free on a shitty laptop or mobile? That's what I'm stuck with but I'd love to mess with it.

You might wanna try google colab/paperspace, there you get access to their GPU cloud and can run SD via webui on their servers, which is super fast. For free you often don't really get access to their stuff and it's highly time-of-day dependant and you'll get kicked off fairly quickly. If you are willing to pay a little, I can't recommend colab pro anymore but paperspace pro for $8/mo. grants you access to high-end and even professional nvidia cards for many, many hours (there's no clear limit, it's highly dependant on how much you use their service and other hidden variables like their load) that'll do the average image generation in seconds instead of hours. There's guides how to set it all up and it's pretty easy. It's worth it IMHO.

 

[Stephanie Bustcakes]

Here's a question for those of you who can actually get this stuff to work; is it possible to use AI art generators to make something like a 3D render in Blender or Daz look more realistic?

Yes.

(click for full res. thumbnails don't do the end result justice.)
Jill Valentine from Resident Evil remake.
original: with img2img:

 

[WhiskeyJack]

Yes.

(click for full res. thumbnails don't do the end result justice.)
Jill Valentine from Resident Evil remake.
original:View attachment 4169706 with img2img:View attachment 4169738

Looks like the AI is a big fan of Alexandra Daddario... This would be very helpful though, thanks. Just getting the hair to look better is a real time saver. That's like two hours of Photoshop work right there.

Like stated above, Img2Img is where you're going to want to focus your efforts. I can guarantee that there were almost certainly zero images of people riding T-Rexs used in the training data of whatever model you're using, so good luck getting a person in a reasonable place on the dinosaur using text prompts. If you have a rough image to work from already that's a better starting point.

Well, with something like Daz I can use existing dinosaur assets to make something like this (not my renders):

And there's a few decent Scarlett Johansson-lookalike models out there

So I can just make the images I need of people (like Black Widow in this instance) riding a T-Rex. However, Daz and Blender renders always end up looking like plastic, unless the lighting is just absolutely perfect, which it never is.

But it sounds like this Img2Img thing would work to enhance things without me spending all day fiddling with lighting after-effects in PS or Gimp. Rendering takes long enough as it is, but it's so versatile for getting a particular scene/pose set up, and still faster than me trying to build everything from scratch in PS.

 

[Catgirls are Love]

But it sounds like this Img2Img thing would work to enhance things without me spending all day fiddling with lighting after-effects in PS or Gimp. Rendering takes long enough as it is, but it's so versatile for getting a particular scene/pose set up, and still faster than me trying to build everything from scratch in PS.

Pretty much. If you want a not dissimilar image from what you give it, Img2Img can do a pretty good job so long as the model has a decent grasp of what it's trying to replicate (or made the source image).

For example, if I take the end result of what I posted a few pages back and run it through Img2Img, this is an example of what I get.

Spoiler: Fantasy Boobplate

Keep in mind that the original image was made via txt2img mode. Now, if I do a quick google search for two guys wrestling with one putting the other in a headlock and run that through img2img...

Spoiler: wat

Results are not quite what was desired, probably partially because of the prompts I gave it but also likely because the model just wasn't trained with sweaty men groping each other. Keep in mind there are also settings to tweak to determine how strictly it mimics the original image.

Spoiler: Model does not know how to draw a Kiwi Bird

Prompts still matter; I can give the model a picture of my cat, put "dog" in the prompt and I'll get dogs as outputs. So you need more than just an image; you also need prompts the model has been trained on to get the result you want. It's still worth a shot, but the more unusual the subject matter the harder it will be to accomplish.