AI Image Generation

Edaw

Parody
<Gold Donor>
12,294
77,854
The prompt says "a woman". I'm curious if there is a built-in bias for what "a woman" is supposed to look like. What percentage is white and/or dark hair, breast size, etc.
Honestly, I think the AI sees things in a very fuzzy way. It probably uses some mashup of all the prompt that leads you to a 1940s pinup look.

"a woman posing next to a stunning vintage car, pastel colors, photograph --ar 16:9"

So, you end up getting a 'posing, stunning, vintage, pastel colors, photograph, woman' and a ''posing, stunning, vintage, pastel colors, photograph, car'.
 

McFly

<Aristocrat╭ರ_•́>
1,311
868
I work with a production team that's embracing AI image-gen for concept artwork. We've been using it to provide animators with concept artwork. It's definitely decreased the amount of time before animators can begin working on various shots.

Just a mere few weeks ago, we were having problems with some historically accurate items, but it seems to be getting better and better, almost by the day. I recall what version 3 of Midjourney was like a year ago and it's amazing how quickly it's progressed.

Even though I'm focused on audio, I can see how this is going to radically change small-scale, artsy, creative productions.

If you're involved in any creative/artistic field, I think you better start at least learning some of the basics about prompting. At least until one of the major companies create the perfect UX/UI for text to image gen.
 
  • 4Like
Reactions: 3 users

Mist

Eeyore Enthusiast
<Gold Donor>
30,411
22,199
I work with a production team that's embracing AI image-gen for concept artwork. We've been using it to provide animators with concept artwork. It's definitely decreased the amount of time before animators can begin working on various shots.

Just a mere few weeks ago, we were having problems with some historically accurate items, but it seems to be getting better and better, almost by the day. I recall what version 3 of Midjourney was like a year ago and it's amazing how quickly it's progressed.

Even though I'm focused on audio, I can see how this is going to radically change small-scale, artsy, creative productions.

If you're involved in any creative/artistic field, I think you better start at least learning some of the basics about prompting. At least until one of the major companies create the perfect UX/UI for text to image gen.
How it's going to change is that everything is going to start looking the same if all the concept artwork is being generated by things that effectively take an average+remix of tons of existing art.

Companies doing this will quickly become indistinguishable from any of the other companies doing this.
 
  • 1Truth!
  • 1Imbecile
Reactions: 1 users

Lambourne

Ahn'Qiraj Raider
2,715
6,531
The prompt says "a woman". I'm curious if there is a built-in bias for what "a woman" is supposed to look like. What percentage is white and/or dark hair, breast size, etc.

If let alone, it tends to go for slender white women in summer dresses which I think is a result of the training set having a wider selection of European/North American photography, most of which was professional models.

This bias will swing another way if there is a cue in the prompt that it associates more heavily with a certain race. The ones below all use the same seed but I specified a different building for the people to be standing in front of and it affected the race of the people as well. It's another bias in the training set, a building from Japan is going to be more heavily associated with Asian faces than one in Mecca. You can even see it in the way it biased towards cold weather clothing for NYC and warm weather clothing for Dubai.

1684572603259.jpeg


1684572616641.jpeg


1684572630489.jpeg


1684572647582.jpeg
 
  • 2Like
Reactions: 1 users

McFly

<Aristocrat╭ರ_•́>
1,311
868
How it's going to change is that everything is going to start looking the same if all the concept artwork is being generated by things that effectively take an average+remix of tons of existing art.

Companies doing this will quickly become indistinguishable from any of the other companies doing this.

I respectfully disagree with this assessment of the direction image gen is going.

I've been using this technology since last July — the ability for variation and differential in the generated artwork has been increasing on every iteration. The ability to "craft" an appropriate prompt with modifiers and creative language has nearly become analogous to a painter choosing the brush and paint.

Collectively, the production team and I have generated thousands of images, and each iteration has been an improvement. The thing we're finding out is that we've had to "re-learn" how to craft statements to achieve a result. It's like a boss fight in a game; once we figured out the correct way to create the prompt, it becomes infinitely easier and more productive. Once we starting using job ID's, seed #'s, blending commands, image/text weights to start reproducing outputs; this became more than a 'party-trick' gimmick. (And this doesn't include using a LLM AI to help out)

I think I can confidently say that if you're a person/company whose focused on both being creative AND being profitable; NOT utilizing this technology is a big mistake.

It's still in it's infancy now, and the speed at which is improving is nearly unfathomable to me. My advice would be to stay as current as you can on the tools and techniques that appear, so you can use it to further an idea or inspiration.

We're at an inflection point in human creativity and I for one am probably the most excited I've been in decades about the potential of this tech. Good times.
 
  • 3Like
Reactions: 2 users

Denamian

Night Janitor
<Nazi Janitors>
7,185
18,963
I've been using this technology since last July — the ability for variation and differential in the generated artwork has been increasing on every iteration. The ability to "craft" an appropriate prompt with modifiers and creative language has nearly become analogous to a painter choosing the brush and paint.

Not to mention the use of different models and LoRAs to get vastly different results using the same prompts.
 

Pasteton

Blackwing Lair Raider
2,602
1,714
So anyone else finding lots of ai generated imagery in their ads? There’s plenty of dumpster quality ads out there and lately a lot more giant hands or extra fingers in ‘em
 

Lambourne

Ahn'Qiraj Raider
2,715
6,531
I don't really get ads but I have noticed an increase in chatbot generated posts on reddit etc, there's a certain wordiness to them that starts to stand out. I think Elon's right in that adding a paid subscription is the only real way to combat massive bot farms taking over Twitter.
 
  • 1Like
Reactions: 1 user

ShakyJake

<Donor>
7,629
19,259
I don't really get ads but I have noticed an increase in chatbot generated posts on reddit etc, there's a certain wordiness to them that starts to stand out. I think Elon's right in that adding a paid subscription is the only real way to combat massive bot farms taking over Twitter.
I've mentioned this elsewhere, but what will happen when Facebook, Twitter, Instagram, etc. etc. are filled with fake accounts? Fake accounts so realistic that users won't be able to tell who is real and who isn't? How will advertisers react when these sites can't accurately state their actual user totals? I think these sites may become useless unless some mechanism is devised to filter the fake users out.
 

Captain Suave

Caesar si viveret, ad remum dareris.
4,768
8,033
I've mentioned this elsewhere, but what will happen when Facebook, Twitter, Instagram, etc. etc. are filled with fake accounts? Fake accounts so realistic that users won't be able to tell who is real and who isn't? How will advertisers react when these sites can't accurately state their actual user totals? I think these sites may become useless unless some mechanism is devised to filter the fake users out.

That's why Elon wants to charge for verification. There isn't going to be a good filter, otherwise.
 

ShakyJake

<Donor>
7,629
19,259
That's why Elon wants to charge for verification. There isn't going to be a good filter, otherwise.
Would you pay for Facebook or Twitter? I wouldn't. Twitter could conceivably have a "read-only" mode for non-paying accounts which I'd be fine with. But Facebook couldn't operate this way.