AI Image Generation

Captain Suave · Jun 23, 2023

ShakyJake said:
Watching various prompts from people on Midjourney, I often see "4K'", "8K", etc. Does this actually make a difference? The resulting images aren't actually that resolution.

Adding to what Mist said, they're using the prompt to pull attributes of training images images which were flagged as 4k/8k, etc. It could be as vague an association as "high-res images are generally produced by more competent artists and are qualitatively better along some metric I care about". Prompts include camera models and such for similar reasons.

Tmac · Jun 25, 2023

Inpainting could be super powerful but the AI has a lot of trouble w it.

It seems like you have to inpaint significant real estate, bc if it’s too small the AI doesn’t know what to do with it.

Lambourne · Jun 25, 2023

New Midjourney function lets you zoom out images, i.e. generate new content around an existing image (presently only MJ generated stuff though). Helpful for optimizing images that looked good but were cropped a little wrong or had the wrong aspect ratio. Even works on images generated earlier.

Base:

Zoomed out:

You can stack this several times, this one was done 3 times:

Mist · Jun 25, 2023

Which one of these is better?

View attachment 479813

View attachment 479814

I like the way the top one looks, but it feels a little too ripped off from Dr. Strange.

Chukzombi · Jun 25, 2023

Mist said:
Which one of these is better?

View attachment 479813

View attachment 479814

I like the way the top one looks, but it feels a little too ripped off from Dr. Strange.

seems odd to have ice particles mixed in with outer fire particles. you could mix in neutral gray or white inner particles and then you can create whatever outer particles.

Captain Suave · Jun 25, 2023

Mist said:
Which one of these is better?

View attachment 479813

View attachment 479814

I like the way the top one looks, but it feels a little too ripped off from Dr. Strange.

I like the first one. Very Dr Strange-ish but I think the color balance is better.

Pasteton · Jun 29, 2023

pretty sick

hopefully i get to make my own virtual world soon. I will release an mmo before pantheon

Kharzette · Jun 29, 2023

I wonder if it could generate clean collision geometry. Those environments would be chaos in a game engine.

Captain Suave · Jul 3, 2023

Valve says Steam games can’t use AI models trained on copyrighted works

“Legal uncertainty” over models means many devs can’t establish “appropriate rights.”…

arstechnica.com

Lambourne · Jul 13, 2023

New Midjourney function lets you expand images left and right, and rather than the earlier zoom out function, this is new content generated at the same resolution. You can also modify the prompt.

For example, I started with a prompt for a fantasy style female thief by a campfire

Then expanded the image but modified the prompt to male knight.

You can redo this of course, so if you wanted to change the male figure you can just go off the base image again but with a different prompt.

Definitely a lot of potential here because you get a lot more direction during the creation process. Before you'd have to just prompt for a male and a female figure by a campfire and hope for the best, now you can get one part of the image right before working on the rest. It's also expanded images not resampled ones so you don't lose any resolution.

Combined, you can create higher resolution images than before. I'm having to scale these down a lot to post them here. The one below is like 2.5mb uncompressed and 2700px wide

Kharzette · Jul 13, 2023

That above looks so good. I love how it uses the campfire light.

I just got a lora to work finally. I'm running about 9 seconds per iteration, so four and a half minutes per 512x512 image. Not sure what I did to make it slower. It feels like it is memory thrashing.

This is the peace sign lora: Still mangles the fingers

16079-417097083-beautiful girl, white babydoll dress, small cute pointy ears, pale skin, short...png

16078-417097082-beautiful girl, white babydoll dress, small cute pointy ears, pale skin, short...png

Kharzette · Jul 13, 2023

Before overwatch league started I ran some experiments with the vtuber-poses lora. This sort of generates what I think the live2D program wants to make a 2D avatar. If you prompt feet, you get a really nice almost orthographic front view.

I think this is what I've been looking for for generating good ortho templates for making 3D characters. If I can get the arms to either V or T. My stuff supports V arms now as I've got capsules working for bone bounds.

Can't post any pictures because I make the templates nude so I can see muscle details and the like.

Kharzette · Jul 13, 2023

Here's some with "catsuit". First the default athletic slender build:

16105-1887667889-beautiful girl, catsuit, small cute pointy ears, standing, pale skin, bald, d...png

A more middle ground. This one the feet weren't emphasized enough so it cut them off:

16108-2633469206-beautiful girl, catsuit, small cute pointy ears, standing, pale skin, bald, d...png

And a couple porky big booba builds:

16114-101384094-beautiful girl, catsuit, small cute pointy ears, standing, pale skin, bald, de...png

16115-101384095-beautiful girl, catsuit, small cute pointy ears, standing, pale skin, bald, de...png

And this one gave me 2 sides. No idea why, haven't been able to reproduce it, but I love it:

16109-2345945730-beautiful girl, catsuit, small cute pointy ears, standing, pale skin, bald, d...png

Rajaah · Jul 13, 2023

Lambourne said:
New Midjourney function lets you expand images left and right, and rather than the earlier zoom out function, this is new content generated at the same resolution. You can also modify the prompt.

For example, I started with a prompt for a fantasy style female thief by a campfire

View attachment 482261

Then expanded the image but modified the prompt to male knight.

View attachment 482262

You can redo this of course, so if you wanted to change the male figure you can just go off the base image again but with a different prompt.

View attachment 482272

Definitely a lot of potential here because you get a lot more direction during the creation process. Before you'd have to just prompt for a male and a female figure by a campfire and hope for the best, now you can get one part of the image right before working on the rest. It's also expanded images not resampled ones so you don't lose any resolution.

Combined, you can create higher resolution images than before. I'm having to scale these down a lot to post them here. The one below is like 2.5mb uncompressed and 2700px wide

View attachment 482263

View attachment 482264

Whoa, those environments are super nice. If this tech had been around 25 years ago it would have been amazing to be able to create stuff like this as a teenager, back when I was obsessed with writing / drawing fantasy stuff. I would have been all in with AI art and it could have given life to a lot of ideas that I wasn't talented enough at drawing to create the way I wanted.

Mist · Jul 13, 2023

I'm from the future. Come with me if you want to be fake and gay.

Kharzette · Jul 15, 2023

Well I gave up on loras. They just seem to have the upper arms pinned to the torso, like that maneuver every young girl learns.

I realized that I had not gotten a single out of memory error all day long. That combined with the slowness that felt like paging and I realized they must have added a memory manager to this thing.

So now when you fail an allocation, it will just find the least recently used chunk and free it, then defrag to get a chunk of the size requested. DirectX does this automatically for textures / normals, but these AI coders have always been too lazy to bother.

So, controlnet here I come! I fired it up and sure enough it runs. The results were terrible, but it does work. Instead of using it directly, which is locked to the standard model with ddim, I'm integrating it with the automatic 1111 ui via a plugin. I've got several gigs of models to download but I shall report back with my findings.

pharmakos · Jul 16, 2023

Been experimenting using album covers or other famous images as input images for Stable Diffusion 2.1. Did low noise for these DSotM remixes.

Mist · Jul 17, 2023

This is what I was talking about in another thread. Surprised it's happening so quickly, though not really. Just a full AI rendering pipeline, fuck polygons.

EDIT: Eh I went to their site, it's just a hack, it doesn't really work. "This demo is a conceptual demonstration of what could soon be the generation of experiences/games in the near future." It's pregenerated textures, generated by stable diffusion, and then stretched over a bunch of AI generated polygons.

pharmakos · Jul 17, 2023

I'd say that's more "proof of concept" than "hack"

Pasteton · Jul 18, 2023

Help me understand what is the roadblock from ai creating ai 3D worlds. Needs more data? Not enough compute ?

AI Image Generation

Caesar si viveret, ad remum dareris.

Adventurer

Ahn'Qiraj Raider

REEEEeyore

Millie's Staff Member

Caesar si viveret, ad remum dareris.

Blackwing Lair Raider

Watcher of Overs

Caesar si viveret, ad remum dareris.

Ahn'Qiraj Raider

Watcher of Overs

Watcher of Overs

Watcher of Overs

Honorable Member

REEEEeyore

Watcher of Overs

soʞɐɯɹɐɥd

REEEEeyore

soʞɐɯɹɐɥd

Blackwing Lair Raider