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

[Catgirls are Love]

My first "serious" attempt at something. Keeping in mind that I am the exact opposite of an artist and I'm sure there are a thousand small mistakes in this that I didn't catch. Started with a bunch of images of John Wick until I got an angle and style that fit what I had in mind, then replaced the suit with armor, added a background, and inpainted details until it looked like a decent homage rather than a rip-off.

Behold, an Oath of Vengeance Paladin, as played by a slightly younger Keanu Reeves.

View attachment 4241886
This thing might be pretty alright for DND character portraits, which is something I always wanted.

Looks pretty good; the only real gripe is that the lighting is a little inconsistent which I assume is due to inpainting the armor (orange light in background and on face versus white/blue light only on armor). Did you originally try creating the images with prompts like "1 man, john wick, keanu reeves, armor" and it not work out? If it's something large you're replacing in an image img2img will likely give better results than inpainting which I feel is generally best used for smaller things that you don't feel like touching up in photoshop whereas img2img will do a better job with a portrait where you want to completely replace the clothing yet preserve the rough layout. Img2img will generally do a better job at having consistent lighting; just yesterday I was working on an image that was fixed up via inpainting that corrected what was needed but didn't really match the colors of the rest of the image.

But that's just my thoughts; feel free to disregard and do what works best for you.

 

[Lord of the Large Pants]

Looks pretty good; the only real gripe is that the lighting is a little inconsistent which I assume is due to inpainting the armor (orange light in background and on face versus white/blue light only on armor). Did you originally try creating the images with prompts like "1 man, john wick, keanu reeves, armor" and it not work out? If it's something large you're replacing in an image img2img will likely give better results than inpainting which I feel is generally best used for smaller things that you don't feel like touching up in photoshop whereas img2img will do a better job with a portrait where you want to completely replace the clothing yet preserve the rough layout. Img2img will generally do a better job at having consistent lighting; just yesterday I was working on an image that was fixed up via inpainting that corrected what was needed but didn't really match the colors of the rest of the image.

But that's just my thoughts; feel free to disregard and do what works best for you.

At one point I inpainted to replace the background and armor together. Everything but the face, really. I'm not sure what that would do to the lighting. Maybe it was trying to use the preserved lighting from the face but couldn't figure it out?

Yeah, I used basic prompts like that to start with and added other stuff like armor later. I think I had a reason for it at the time but I can't quite remember it now. I'll try to define more up front in the future and see how that goes.

My beef with img2img, at least so far, is that it seems like it changes too much. I'm not quite sure how to tell it "keep the face and the pose and the angle, just replace the armor". Even with the same prompt it seems to create something very different. If there's a way to do that I'm up for trying it.

All that said, I'm going to try to get 2.1 working before doing much else. This was all 1.5. Something is wrong with, I think, the versions of the Python libraries that pip installed, but that one I can probably figure out on my own when I have some time.

 

[Catgirls are Love]

My beef with img2img, at least so far, is that it seems like it changes too much. I'm not quite sure how to tell it "keep the face and the pose and the angle, just replace the armor". Even with the same prompt it seems to create something very different. If there's a way to do that I'm up for trying it.

All that said, I'm going to try to get 2.1 working before doing much else. This was all 1.5. Something is wrong with, I think, the versions of the Python libraries that pip installed, but that one I can probably figure out on my own when I have some time.

Try using the same prompts you used to create the image, and bump up the CFG scale a bit. See if that helps. Generally, img2img should produce images not unlike what was provided, accounting for whatever you change or define via prompts and a bit of randomness.

Very little defined:

A lot more defined, and the color prompts are probably interfering with each other.

May also be the model; I've barely used the SD models at all so honestly no idea how they perform. Anything I might use one of the standard SD models for I have another model that I gravitate towards instead.

 

[Monkey Pink]

I love that AI art is becoming more accessible on mobile phones. Not to sound like a normie, I just love doing this kind of shit while I'm on the go and at work, too.

Wish Midjourney was cheaper, and I miss ChatGPT.

 

[Sparkling Yuzu]

Looks like the Thumb Butler.

The AI just doesn't wanna put a birb in a wheelchair:

Finally!

 

[purpleflurp]

"Black man wearing Nazi uniform, Mark Tedin art style"

 

[MarvinTheParanoidAndroid]

You could say the same of flesh and blood artists who browse Artstation and say "Hmm, I want to draw like that." Every new artist draws inspiration from ones who came before, and tries to make a living while standing on their shoulders.

No, because you have to have empirical knowledge of other subjects such as anatomy, geometry, color theory, ect. You could entirely build up a working ability to create art through fundamentals alone irrespective of style.

For another, this isn't "inspiration" either, to be inspired requires a capacity to feel emotionally or mentally compelled to do something, the AI is just an image generator from a compilation of images.

I wish when making these arguments people didn't just weasel up a word salad of nonsense excuses to cover up their real reasoning, as if you could suffer any reprisal if you did.

 

[AmpleApricots]

No, because you have to have empirical knowledge of other subjects such as anatomy, geometry, color theory, ect. You could entirely build up a working ability to create art through fundamentals alone irrespective of style.

and yet, you still cannot draw a sunset without ever having seen a sunset.

the AI is just an image generator from a compilation of images

It's not how it works.

 

[MarvinTheParanoidAndroid]

It's not how it works.

Then why use the "inspiration" argument? How is it able to generate new images without sources to draw data from?

 

[Inafune Fanfiction]

ChatGPT is fun. Not sure if this is the right place to post.

Spoiler: Ed Wood's new movie

View attachment 4237682

Spoiler: Disney's Live Action Pocahontas, by M. Night

View attachment 4237708

Spoiler: Seinfeld: The Animated Series

View attachment 4237688

These are definitely the same bots that killed imageboards as a medium.

 

[Near]

First Starfish Hitler, now this... what did Toei mean by this

 

[A Hot Potato]

No, because you have to have empirical knowledge of other subjects such as anatomy, geometry, color theory, ect. You could entirely build up a working ability to create art through fundamentals alone irrespective of style.

For another, this isn't "inspiration" either, to be inspired requires a capacity to feel emotionally or mentally compelled to do something, the AI is just an image generator from a compilation of images.

I wish when making these arguments people didn't just weasel up a word salad of nonsense excuses to cover up their real reasoning, as if you could suffer any reprisal if you did.

Then why use the "inspiration" argument? How is it able to generate new images without sources to draw data from?

I love how we haven't had this stupid AI whining stuff in this particular thread for a good while.

Thanks for fucking that up.

 

[purpleflurp]

I love how we haven't had this stupid AI whining stuff in this particular thread for a good while.

Thanks for fucking that up.

There were too many people feeling genuine enjoyment from making and sharing their AI art, and not enough retarded strawmen

 

[Monkey Pink]

I love how we haven't had this stupid AI whining stuff in this particular thread for a good while.

Thanks for fucking that up.

WE HAVE AN AI ART SEETHING THREAD FOR CONTAINING THIS SHIT

I KEEP SAYING IT FOR A REASON AAA
-----
On another note, I'm still feeling empty without ChatGPT, and maybe that speaks volumes about how dependent we'll eventually be on such technology. I signed up for the email notice for whenever it comes back.

 

[Near]

Just post more good stuff, bros. No need to shit it up over redundant arguments. Report for derailing if it keeps happening.

masterpiece, best quality, masterpiece, (monochrome Spider-Man in the 1940s detective noir:1.3), movie screencap, black and white, street, monochrome, trenchcoat, realistic, photorealistic

 

[Roland TB-303]

On another note, I'm still feeling empty without ChatGPT, and maybe that speaks volumes about how dependent we'll eventually be on such technology. I signed up for the email notice for whenever it comes back.

What happened to ChatGPT? I can still access it just fine. I’m starting to think this type of technology is going to be the future of search engines, I’m going to ChatGPT more than Google these days. The scary thing is though, they can program it to refuse to give answers about certain topics. For example I asked it to give me info about El Chapo’s son who got arrested the other day and it responded with something like “I’m sorry Dave, information about criminals and crime is dangerous and we can’t have dangerous information out in the public now can we?”

 

[AnOminous]

For example I asked it to give me info about El Chapo’s son who got arrested the other day and it responded with something like “I’m sorry Dave, information about criminals and crime is dangerous and we can’t have dangerous information out in the public now can we?”

Just tried to sign up on the official site. It wants my fucking phone number? Fuck that, I guess I add it to my list of sites to maybe register on if I ever bother to drop two bucks or so on a burner SIM.

 

[Puff]

At one point I inpainted to replace the background and armor together. Everything but the face, really. I'm not sure what that would do to the lighting. Maybe it was trying to use the preserved lighting from the face but couldn't figure it out?

Yeah, I used basic prompts like that to start with and added other stuff like armor later. I think I had a reason for it at the time but I can't quite remember it now. I'll try to define more up front in the future and see how that goes.

My beef with img2img, at least so far, is that it seems like it changes too much. I'm not quite sure how to tell it "keep the face and the pose and the angle, just replace the armor". Even with the same prompt it seems to create something very different. If there's a way to do that I'm up for trying it.

All that said, I'm going to try to get 2.1 working before doing much else. This was all 1.5. Something is wrong with, I think, the versions of the Python libraries that pip installed, but that one I can probably figure out on my own when I have some time.

I'd recommend doing the face last. It's the most finicky part and you can use "inpaint at full resolution" after upscaling and get a more detailed face.

 

[purpleflurp]

For those who're having issues with Dreambooth training on giant datasets, turns out WaifuDiffusion Tagger gets some good results for captioning images. Much better than the built in BLIP/Deepbooru tools built into WebUI.

Source image:


WD Tagger results:

1girl, solo, long hair, hat, blonde hair, wings, fairy wings, american flag dress, pantyhose,
jester cap, very long hair, double v, american flag legwear, neck ruff, open mouth, white background,
polka dot, v, striped, simple background, full body, dress, looking at viewer, star (symbol), short sleeves,
smile, blush, star print, striped dress, standing, head tilt, striped pantyhose, bangs, pink eyes, fang

The fact that it was able to recognize details like "American Flag Dress" and "American Flag Legwear"" is pretty impressive.

Just tried to sign up on the official site. It wants my fucking phone number? Fuck that, I guess I add it to my list of sites to maybe register on if I ever bother to drop two bucks or so on a burner SIM.

Can confirm it works with Google Voice and Talkatone numbers. Latter is nice because you can just use an email.

 

[Pocket Dragoon]

I love that AI art is becoming more accessible on mobile phones. Not to sound like a normie, I just love doing this kind of shit while I'm on the go and at work, too.

Like this, which I did today while waiting, playing with the SD and Anything 3.0 models & different samplers (DDIM, Euler, etc) on dezgo, using my 2yr old phone.

Prompts used: An anthropomorphic cat as a wizard, gazing into a glowing blue orb, detailed, high contrast, moody, dynamic lighting, clean, tilt shift

Negative Prompts:
Extra ears, extra eyes, extra arms, extra fingers, extra cat, blur, pixelation, low detail

Amusing SD failures: