|
Here's "Carl Sagan as an Apex Legends character"
|
#
?
Aug 9, 2022 19:23
|
|
|
|
| # ? May 30, 2024 07:03 |
|
|
Keep our Guy rolling...     And a samurai mouse having tea 
|
|
#
?
Aug 9, 2022 20:27
|
|
|
Guy Fieri action figure, product photo, detailed, 4k     
|
|
#
?
Aug 9, 2022 20:41
|
|
|
Condoleezza Rice Lego Minfig   Action figures  Commemorative plate  Skyrim  Resident Evil  Super Smash Bros  On Star Trek the original series  Younger and in Star Wars  Some sort of anime prompt I stole idk
|
|
#
?
Aug 9, 2022 20:57
|
|
|
Got into the Dall-E 2 beta, and.... I think their promo images are cherry picked as gently caress. Only been messing around with this for half an hour, but I'm rating this below Midjourney and Stable Fusion. It can turn out some amazing stuff occasionally, but the ratio of good to crap seems way lower than the other two programs.
|
|
#
?
Aug 10, 2022 00:13
|
|
|
Bloody Hedgehog posted:Got into the Dall-E 2 beta, and.... I think their promo images are cherry picked as gently caress. Only been messing around with this for half an hour, but I'm rating this below Midjourney and Stable Fusion. It can turn out some amazing stuff occasionally, but the ratio of good to crap seems way lower than the other two programs. I think Dall-E 2 seems to have a worse language model than all the other systems, it's so bad at understanding prompts. Which is odd considering they also made GPT-3. In google's write-up about Imagen they noted that improvements to their language model made more impact than further training on their visual models. I really hope that become's a large focus for all these team's going forward, would be lovely if they all started understanding complex prompts. Could make even weirder stuff.
|
|
#
?
Aug 10, 2022 00:20
|
|
|
Bloody Hedgehog posted:Got into the Dall-E 2 beta, and.... I think their promo images are cherry picked as gently caress. Only been messing around with this for half an hour, but I'm rating this below Midjourney and Stable Fusion. It can turn out some amazing stuff occasionally, but the ratio of good to crap seems way lower than the other two programs. depends on what you're trying to do, the discord has some good help on prompts. Personally I've had a lot of trouble generating consistent numbers of arms and legs on midjourney and SD, while that hasn't been a big issue on dalle. Tunicate fucked around with this message at 02:59 on Aug 10, 2022 |
|
#
?
Aug 10, 2022 00:58
|
|
|
It's got no idea what the shaka brah in "shaka brah wearing oakleys" is but whatever, those look cooler than any oakleys Yeah from my experience so far Midjourney is better than DALL-E is better than GAN+CLIP, but the first two are both miles beyond G+C See More Butts fucked around with this message at 01:23 on Aug 10, 2022 |
|
#
?
Aug 10, 2022 01:18
|
|
|
I've had a bunch of prompts that generated mediocre results but the AI sure does love generating angry animals doing paperwork:  
|
|
#
?
Aug 10, 2022 03:23
|
|
|
Mercury_Storm posted:"inside a dystopian and post apocalyptic Walmart where swarms of rats cover the floor and smiling families trade ammunition for food, in the style of H.R. Giger" -n 9 -g Were these done by Craiyon? I've found that it's occasionally able to produce delightfully creepy stuff beyond just its inability to make good faces. edit: Has anyone come up with good small-scale applications for these images via creative commons? I've been messing around with makeplayingcards.com to try and design card backs for a board/card game.  The "low resolution" tag is a result of the image not meeting a minimum resolution requirement, which it says will result in a blurry or pixelated printing. Anyone know if just upscaling could help with that? I'll report back once I get my "fast" hours back and can upscale to max. Ceramic Shot fucked around with this message at 03:48 on Aug 10, 2022 |
|
#
?
Aug 10, 2022 03:35
|
|
|
Ive bee using stable fusion for everything for the past couple days, you can generate 9 images like craiyon does using the -n 9 -g arguments. Here's a "dystopian ratty Lovecrafian 2020 Doritos television commercial still showing a smiling family in living room sacrificing a chicken the style of H.R. Giger"  "Bob Ross performing a finishing move in Mortal Kombat"  "Bob Ross holding a baby seal" 
Mercury_Storm fucked around with this message at 15:10 on Aug 10, 2022 |
|
#
?
Aug 10, 2022 03:49
|
|
|
Mercury_Storm posted:"Bob Ross performing a finishing move in Mortal Kombat" Uppercutting the opponent into the stratosphere and then painting a jagged mountain formation in the distance that his victim gorily splats into, just in time to finish the phthalo blue skyline and call your life a happy little accident.
|
|
#
?
Aug 10, 2022 16:14
|
|
|
How does stable diffusion handle the prompt "a horse riding an astronaut"? Every other model desperately wants to make the astronaut ride the horse.
|
|
#
?
Aug 10, 2022 18:30
|
|
|
Wheany posted:How does stable diffusion handle the prompt "a horse riding an astronaut"? 
|
|
#
?
Aug 10, 2022 18:39
|
|
|
Wheany posted:How does stable diffusion handle the prompt "a horse riding an astronaut"? was worried this would get gross since there's no filters at all and you can totally generate porn easily. (I've seen some Goblin dong recently with no mention of anything about nudity or anything beyond just wanting a goblin doing something) Anyway same thing, I don't think prompt order matters a ton with this yet. SD really doesn't like multiple subjects right now and they seem to mostly be interested in letting you make larger works and have the AI work on 1 character at a time rather than print an entire image. So you'd like get a horse then you'd get the horse out then you'd get your astronaut then you'd place them then you'd ask it to do the combined and it would do that. From the voice chat meetings on discord that sounds like the direction they want this to go. 
|
|
#
?
Aug 10, 2022 18:42
|
|
|
pixaal posted:I've seen some Goblin dong recently Hehe, wow that's wacky. You, uh, probably wouldn't want to post that though unless...?
|
|
#
?
Aug 10, 2022 19:04
|
|
|
dot communist posted:Hehe, wow that's wacky. Uh you probably wouldn't want to post that though I didn't save it, and I reported the generation for ToS violation so I wouldn't get removed from beta, it's gone If you want that, they are going to release a notebook, you'll be able to run that on your RTX 3000 series card yourself I doubt an RTX is doing this is 4seconds like whatever the hell they are running this on.
|
|
#
?
Aug 10, 2022 19:04
|
|
|
you should have saved the goblin dong. these are unique works of art, that's like throwing out a new Rembrandt painting
|
|
#
?
Aug 10, 2022 19:12
|
|
|
Rutibex posted:you should have saved the goblin dong. these are unique works of art, that's like throwing out a new Rembrandt painting I don't have the space for that! SD lead developer posted:Phew what a few days eh. I get the sense the devs just want this to exist and be public access and don't care what you want to do with it but don't gently caress with Discord ToS because that's how they are running this and they don't want to get kicked off discord. They are enforcing it. pixaal fucked around with this message at 19:39 on Aug 10, 2022 |
|
#
?
Aug 10, 2022 19:36
|
|
|
I got dalle2 access! anthro lion and anthro tiger is arm wrestling in a dimly lit bar, digital art, trending on artstation 4k  action scene from a John Wick movie as claymation  Panda made of black and white melted chocolate on a dessert plate, Michelin star restaurant, 4K photo  Mewtwo statue made by Jeff Koons  Zoidberg statue made by Jeff Koons  Homer Simpson statue made by Jeff Koons I don't get what's up with Dalle2 and pop culture references - Mewtwo and Zoidberg looks like cheap bootleg knockoffs, but Homer is like picture perfect in all generations some more of my generations: https://imgur.com/gallery/3hrO9L5
|
|
#
?
Aug 10, 2022 19:38
|
|
|
I'm also interested in the goblin dong. Just saying
|
|
#
?
Aug 10, 2022 19:43
|
|
|
ymgve posted:
Dalle 2 is specifically trying to filter out and prevent you from using names that fall under copyright.
|
|
#
?
Aug 10, 2022 19:43
|
|
|
Is there a beta signup for DALLE 2 somewhere?
|
|
#
?
Aug 10, 2022 19:44
|
|
|
pixaal posted:Dalle 2 is specifically trying to filter out and prevent you from using names that fall under copyright. That's not what's happening, because you can clearly see they resemble what I want, in color scheme if nothing else. The words aren't filtered, so it has to be something at the training step. And that Homer isn't affected at all makes it even weirder Mercury_Storm posted:Is there a beta signup for DALLE 2 somewhere? Sign up for the waitlist at https://labs.openai.com/waitlist - it took me about one month to get access, but they are adding people much more quickly now since they've started taking payment beyond the first 50 generations Also, be warned that they are really strict about certain words - I got warned because I used "burning" in one prompt and "dead" in another. (It wasn't even a reference to dead people or animals, it was the opening line of Neuromancer!) ymgve fucked around with this message at 19:50 on Aug 10, 2022 |
|
#
?
Aug 10, 2022 19:45
|
|
|
AI Generated Art Station: I'm Also Interested In the Goblin Dong
|
|
#
?
Aug 10, 2022 20:02
|
|
|
ymgve posted:That's not what's happening, because you can clearly see they resemble what I want, in color scheme if nothing else. The words aren't filtered, so it has to be something at the training step. And that Homer isn't affected at all makes it even weirder No it's exactly what is happening, Dall-E is altering your prompts before you get your image. The data is in the model, they are intentionally and deliberately not giving you what you ask for. Up until recently they had it against the rules to generate realistic looking people at all, they are exceptionally conservative with what they allow you to see.
|
|
#
?
Aug 10, 2022 20:28
|
|
|
KakerMix posted:No it's exactly what is happening, Dall-E is altering your prompts before you get your image. The data is in the model, they are intentionally and deliberately not giving you what you ask for. So it transformed "zoidberg" into "red anthropomorphic crustacean man" or something before passing that on to the ai?
|
|
#
?
Aug 10, 2022 20:31
|
|
|
KakerMix posted:No it's exactly what is happening, Dall-E is altering your prompts before you get your image. The data is in the model, they are intentionally and deliberately not giving you what you ask for. Got any source of this? I know they are changing prompts to create extra diversity, but I've not heard of them changing prompts to avoid copyright/trademark issues. Besides, how would they do it? The general shape of Zoidberg and his color is still there, how are they changing the prompt to preserve that but not the name of the character?
|
|
#
?
Aug 10, 2022 20:32
|
|
|
I think there's just more homer than zoidberg in the training data you see this in all the systems, even stable diffusion, there are some people that get replicated perfectly and others where it seems like they're going off a rough description
|
|
#
?
Aug 10, 2022 20:42
|
|
|
I guess I would imagine the model doing a basic Google image search, at which point, the hundreds of thousands of images of zoidberg that look like zoidberg would inform the output. It's not like the prompt was for something really obscure, like a third string Dickens character from Bleak House.
|
|
#
?
Aug 10, 2022 21:02
|
|
|
A Strange Aeon posted:I guess I would imagine the model doing a basic Google image search, at which point, the hundreds of thousands of images of zoidberg that look like zoidberg would inform the output. The AI models don't Google anything, they're trained on a labeled data set. Stable diffusion uses this one: https://laion.ai/blog/laion-5b/ There are lots of articles and papers describing in plain english exactly how the various image AIs work.
|
|
#
?
Aug 10, 2022 21:16
|
|
|
Next prompt: Double mermaid, a legendary creature with the top half of a fish and the bottom half of a fish. Referencing this: 
|
|
#
?
Aug 10, 2022 21:22
|
|
|
Midjourney seems to yield better results than dall-e 'darth guy feiri'  'darth john candy'  'undead darth biden'  'undead shogun of harlem'   'yoda with a purple afro' 
|
|
#
?
Aug 10, 2022 23:17
|
|
|
"John candy as an evil dark jedi, painterly style" just your prompt only returned darth vader helmets in SD
|
|
#
?
Aug 10, 2022 23:27
|
|
|
I omitted 'detailed, high quality' from my prompt, which I have used in everything I posted 'spongebob squarepants in the style of wayne barlowe, detailed, HQ'   [ 
|
|
#
?
Aug 10, 2022 23:43
|
|
|
The sickest mosh pit you've ever seen 
|
|
#
?
Aug 10, 2022 23:49
|
|
|
The Atlantic uses Midjourney to produce an illustration for an article https://twitter.com/adamszym/status/1557497131044331526
|
|
#
?
Aug 11, 2022 01:25
|
|
|
I did some copycat generations of the "rap album cover featuring a moai statue" prompt.    And then I tried a power metal album cover featuring a moai statue  
|
|
#
?
Aug 11, 2022 02:31
|
|
|
RAP II is my favorite album
|
|
#
?
Aug 11, 2022 02:36
|
|
|
|
| # ? May 30, 2024 07:03 |
|
|
Not very "album" like, but I do dig the blinged out rabbits. The original prompt was for a blinged out tank, since I wanted to use AI to recreate No Limit album covers, but it never added a tank.
|
|
#
?
Aug 11, 2022 03:04
|
|
