|
Scholtz posted:I thought the celebrity aspect would trigger ToS. The limit is arbitrary and impossible to guess, much like their attempt at fixing racial bias by randomly adding words like "asian" or "black" into your prompt it seems like they did this in laziest and most asinine way possible. I have had so many innocent prompts blocked for violating TOS and the worst part is they won't even tell you which part they think you violated. I tried to make a picture of a boxer shadowboxing and I got a warning. Then I tried a different prompt without the word boxer and it was ok.  Still not sure if making pictures of boxers is blocked or if it was something else about the construction of my prompt. I can't wait for craiyon or some other group to come out with a free or cheap unfiltered dall-e 2 equivalent. It will be so much more fun than counting tokens and trying to fight with openAIs terrible filtering and attempts at bias adjustment. I wouldn't be surprised if such a thing appeared in the next year, craiyon is already pretty comparable to the first Dall-e released in 2021, just did some comparisons against the original demo images they released: (note: Dall-e has an advantage here as they got to massage the prompts for their system and cherry pick the things that worked well)          
|
#
?
Jul 24, 2022 20:17
|
|
|
|
| # ? May 28, 2024 19:07 |
|
|
Scholtz posted:I thought the celebrity aspect would trigger ToS. Nah "Charlie Kelly" is a fictional character and if I use the name on its own it seems to be fine, although what it generates doesn't look quite like Charlie Day 
KinkyJohn fucked around with this message at 21:13 on Jul 24, 2022 |
|
#
?
Jul 24, 2022 21:09
|
|
|
Wheany posted:I saw a post that said had to massage the prompt a bit because the results want so badly to put a deformed elon musk face and nothing else, but Craiyon gives no fucks  
|
|
#
?
Jul 24, 2022 21:12
|
|
|
 "officer down"
|
|
#
?
Jul 24, 2022 21:28
|
|
|
Mercury_Storm posted:had to massage the prompt a bit because the results want so badly to put a deformed elon musk face and nothing else, but Craiyon gives no fucks I like to think the top middle is on stage during one of his bullshit presentations
|
|
#
?
Jul 24, 2022 21:33
|
|
|
ymgve posted:Click the image and use the report function in the lower left - the report function is also for nonsense generations, not just offensive stuff I applied way back in April. I think after leaving closed alpha or beta or whatever they just opened up all the floodgates of people wanting in on the closed beta and let everyone just go broke spending credits on random poo poo they thought of while high instead.
|
|
#
?
Jul 24, 2022 21:46
|
|
|
Talked with one of the developers, the filter on DALLE is a combination of a wordlist ban and some sort of machine learning classifier (I'm guessing GPT3 since they have it in house). After noticing that virtually 100% of generations using the phrase were trying to make porn, they banned 'furry' and 'fursona'.
|
|
#
?
Jul 24, 2022 21:58
|
|
|
 This prompt worked pretty well. Someone care to try it on Dall-e or Midjourney?Frazetts is a typo s/Frazetta.
|
|
#
?
Jul 24, 2022 22:03
|
|
|
Tunicate posted:Talked with one of the developers, the filter on DALLE is a combination of a wordlist ban and some sort of machine learning classifier (I'm guessing GPT3 since they have it in house). LOL, they should just give the people what they want
|
|
#
?
Jul 24, 2022 22:05
|
|
|
A Strange Aeon posted:LOL, they should just give the people what they want but the danger! the incredible danger of an unfiltered image AI! if they were to do that just imagine the horrorscape, an internet filled with furry porn drawings! glad we don't have that problem
|
|
#
?
Jul 24, 2022 22:09
|
|
|
ProfessorMarvel posted:I applied way back in April. I think after leaving closed alpha or beta or whatever they just opened up all the floodgates of people wanting in on the closed beta and let everyone just go broke spending credits on random poo poo they thought of while high instead. I applied at the start of June, so not that long ago, will probably take weeks before I get an invite then Tunicate posted:After noticing that virtually 100% of generations using the phrase were trying to make porn, they banned 'furry' and 'fursona'. Shocked that they didn't think about this before the first invite, does rule 34 work? edit: Dalle Mini definitely doesn't know what rule 34 is 
ymgve fucked around with this message at 22:30 on Jul 24, 2022 |
|
#
?
Jul 24, 2022 22:11
|
|
|
Thinking about it, are there actual ways to tell if an image was ai generated or 'real'? Like is there some secret watermark that would always be present or something? Just trying to imagine what weight photographic evidence or even video will carry in the future, if ai can generate such things convincingly.
|
|
#
?
Jul 24, 2022 22:43
|
|
|
A Strange Aeon posted:Thinking about it, are there actual ways to tell if an image was ai generated or 'real'? Like is there some secret watermark that would always be present or something? If there is some secret watermark in the Dall-E 2 images, wouldn't telling us about it defeat its purpose? There are projects to train AIs in classifying images as real or generated, so I guess that's the way to go.
|
|
#
?
Jul 24, 2022 22:48
|
|
|
A Strange Aeon posted:Thinking about it, are there actual ways to tell if an image was ai generated or 'real'? Like is there some secret watermark that would always be present or something? Thought I saw something about AIs being trained to recognize AI-generated images but that was before they were super convincing. Pictures, video and audio are already well on their way out as evidence when it's presented without context.
|
|
#
?
Jul 24, 2022 22:50
|
|
|
A Strange Aeon posted:Thinking about it, are there actual ways to tell if an image was ai generated or 'real'? Like is there some secret watermark that would always be present or something? I think most of the current AIs are trained with adversarial networks, so there's one that's generating images and another one that's calling bullshit when it can tell the image isn't real and they're in competition with each other. So openAI probably has some decent AI tools at their disposal that can spot generated images with some degree of accuracy. But this isn't even a new problem, even before computers existed people were coming up with convincing ways of faking images. And if you're talking evidence in a courtroom, well, what they consider to be valid evidence is already problematic and full of bullshit (fingerprint analysis is non-scientific nonsense and you'll get different results from different "experts" with the same prints).
|
|
#
?
Jul 24, 2022 22:55
|
|
|
Seems like the craiyon model has been vastly improved in the last month, this is what I got from the website This is what I get from my local copy that I patched together at the start of June: 
|
|
#
?
Jul 25, 2022 00:45
|
|
|
ymgve posted:Seems like the craiyon model has been vastly improved in the last month, this is what I got from the website Can I ask what you're running it on?
|
|
#
?
Jul 25, 2022 00:50
|
|
|
Rinkles posted:Can I ask what you're running it on? 3080ti, it takes about 10-11 seconds to generate each image
|
|
#
?
Jul 25, 2022 00:52
|
|
|
I was wondering if I could squeeze it into my 3060ti.
|
|
#
?
Jul 25, 2022 01:00
|
|
|
Rinkles posted:I was wondering if I could squeeze it into my 3060ti. If it's the 12GB model, you can run the "mega" model, but if you only have 6GB you can probably run the "mini" model or you can use the colab notebook and use google's servers and not worry about your GPU or RAM https://colab.research.google.com/github/borisdayma/dalle-mini/blob/main/tools/inference/inference_pipeline.ipynb ymgve fucked around with this message at 01:08 on Jul 25, 2022 |
|
#
?
Jul 25, 2022 01:03
|
|
|
ymgve posted:If it's the 12GB model, you can run the "mega" model, but if you only have 6GB you can probably run the "mini" model The ti has 8 gigs. How long did it take you to set up?
|
|
#
?
Jul 25, 2022 01:10
|
|
|
Rinkles posted:The ti has 8 gigs. How long did it take you to set up? Way too long, the main issue was creating a proper version of jaxlib that worked on Windows - if that's your platform I got a precompiled jaxlib .whl if you need it.
|
|
#
?
Jul 25, 2022 01:16
|
|
|
ymgve posted:Way too long, the main issue was creating a proper version of jaxlib that worked on Windows - if that's your platform I got a precompiled jaxlib .whl if you need it. Thanks, but I would first need to figure out what I'm doing.
|
|
#
?
Jul 25, 2022 01:30
|
|
|
Rinkles posted:Thanks, but I would first need to figure out what I'm doing. If you just want to play on your Windows machine, I made a full package that's just "click the .bat file to run", but that's the older model so if you want good results it's better to just use the craiyon website ymgve posted:I've been trying to make a "just download and run" package for Dall-E Mini - I haven't tested in on a fresh Windows install, but it should in theory only need the Cuda 11.7 toolkit, then just run the .bat file:
|
|
#
?
Jul 25, 2022 01:36
|
|
|
thanks, that's really cool
|
|
#
?
Jul 25, 2022 01:44
|
|
|
a man covered in mayonaise  https://www.youtube.com/watch?v=mRntutn8udw   Pete Davidson lookin' motherfucker.
|
|
#
?
Jul 25, 2022 06:23
|
|
|
sporklift posted:a man covered in mayonaise wow goddamn
|
|
#
?
Jul 25, 2022 15:18
|
|
|
sporklift posted:a man covered in mayonaise Goddamn, the porn industry has a really tough decade ahead of it
|
|
#
?
Jul 25, 2022 15:56
|
|
|
MidJourney just updated their algorithm to version 3, and it makes much better faces now. Here's some mayonaise covered men  Several of them at the same time: 
|
|
#
?
Jul 25, 2022 22:46
|
|
|
used up a dall-e 2 credit replicating this post from another thread unfortunately meth is against their terms of service so I did "Tooth fairy giving out baggies of white powder, painting by Norman Rockwell"    
|
|
#
?
Jul 25, 2022 22:58
|
|
|
These are all incredible, lmao! Someone please try mustard man, if it knows. Also can DALL-E2 upscale your own image? I think someone mentioned it working to insert objects or expand an image, and upscaling should be a fundamental part of the algorithm I think 
|
|
#
?
Jul 25, 2022 23:01
|
|
|
sadly upscaling is not currently a dalle feature
|
|
#
?
Jul 26, 2022 02:13
|
|
|
Not mine, but I felt this came out surprisingly, disgustingly well. 
|
|
#
?
Jul 26, 2022 13:52
|
|
|
 I like the way this one turned out. A giant uterine monster pouring blood all over the landscape.
|
|
#
?
Jul 26, 2022 14:08
|
|
|
|
|
#
?
Jul 26, 2022 17:16
|
|
|
|
|
#
?
Jul 26, 2022 22:04
|
|
|
Got my DALL-E 2 invite today! I signed up in late April. An astronaut free-floating in space and looking over the Earth as he plays a beautiful melody on an electric guitar, in the style of a psychedelic 1970s concert poster  (poo poo I might actually use this one for my band) Frodo Baggins haggles with a pawn shop owner and tries to sell him the One Ring, oil painting in the style of John Frederick Lewis  A beautiful expressive painting of a burning MBTA train on a bridge over the Charles River with the Boston skyline in the background   
bees everywhere fucked around with this message at 23:22 on Jul 26, 2022 |
|
#
?
Jul 26, 2022 22:58
|
|
|
Just been loving around and some of them come out great but it seems like 2/3 are a mile a way from what I was hoping for. gonna have to look up some tips to get the best results cause I'm wasting a lot of them already. They are gonna get my 15 bux probably lol. Some good ones though A grainy photo of a monster made of pure evil lurking in dark woods   A black and white photo of a demon standing in the corner of a dark bedroom  
|
|
#
?
Jul 26, 2022 23:52
|
|
|
https://youtu.be/XNPMYRlvySY
|
|
#
?
Jul 27, 2022 00:08
|
|
|
|
| # ? May 28, 2024 19:07 |
|
|
haha nice. Does Dall E 2 have a bunch of filters to block out licensed stuff that Dall E mini does not? Seems to impossible to generate anything using real characters for me. Like, I tried to make multiple funny robocop related ones and it worked in mini, but Dall E 2 just generates random robots.
|
|
#
?
Jul 27, 2022 00:22
|
|