|
Patrick Stewart with a penis for a thumb giving a thumbs up? e: new page blah
|
#
?
Oct 20, 2022 23:54
|
|
|
|
| # ? May 29, 2024 15:00 |
|
|
pixaal posted:Patrick Stewart with a penis for a thumb giving a thumbs up? The prompt was literally just "captain picard".
|
|
#
?
Oct 20, 2022 23:55
|
|
|
KPC_Mammon posted:What kinds of prompts do you use to get such vividly colorful output? I've seen people put ((muted colors)) into negative prompts, though if it does anything I don't know
|
|
#
?
Oct 21, 2022 00:00
|
|
|
Tried to prompt a non-cutesy bat for Dalle2's promptober and it straight-up gave one a dick and balls? Then I got a cooler one that just needed some wing tips and feet, except their outpainting sucks poo poo and kept doing dumb stuff like this  
|
|
#
?
Oct 21, 2022 00:01
|
|
|
Rahu posted:Stability AI has issued a DMCA takedown against the Runway ML repo with the 1.5 model, grab it while you can. Anyone got a link to it? WhiteHowler posted:Anyway, here's another thing I made today that I really like. The last image I posted took three hours, over a dozen inpainting sessions, and 600+ iterations. This one took one minor prompt edit, four candidate outputs, one refinement with a higher step count, and a single inpainting session. Where's the prompt, bucko? So what build are people using instead of the Anonymous1111?
|
|
#
?
Oct 21, 2022 00:49
|
|
|
Neat space horse.
|
|
#
?
Oct 21, 2022 01:31
|
|
|
Fuzz posted:Where's the prompt, bucko? "A woman made of the universe, professional painting, beautiful, magnificent" Negative: ((distorted)), deformed, signature, watermark Her original face wasn't terrible but didn't fit the mood. I like the end result; her expression could the wonder of an eager explorer or the satisfied confidence of a creator-goddess. Now that I think about it, there were two inpainting sessions. Her hair was a little bit too stylized, so I did some inpainting touch-up to make it flow better with the ethereal feel of the piece.
|
|
#
?
Oct 21, 2022 02:05
|
|
|
Sedgr posted:Neat space horse. Very cool! I did something similar a few days ago.  "a woman holding a sci-fi large laser gun is riding a ((robot horse)), splash art, detailed faces, light dust, magnificent, digital painting" The only inpaint is her laser gun, it was a nightmare getting her to hold anything that looked halfway decent.
|
|
#
?
Oct 21, 2022 02:09
|
|
|
And not quite the Pegasus I asked for but cool none the less.
|
|
#
?
Oct 21, 2022 02:49
|
|
|
Some more space. Even ones that dont entirely work are still interesting sometimes. 
|
|
#
?
Oct 21, 2022 04:11
|
|
|
Stability.ai's CIO made a blog post and answered some questions on Reddit. They maintain that 1.5 was leaked by RunwayML, which they call "a bad faith actor" who reneged on promises made privately. Someone from Stability.ai did request HuggingFace to take down the model, but Emad on discord clarified that it was not a "legal takedown request". Later they dropped the takedown request altogether and it seems that the leaked model can stay up. https://danieljeffries.substack.com/p/why-the-future-of-open-source-ai 
|
|
#
?
Oct 21, 2022 04:39
|
|
|
Tunicate posted:Its supposed to filter all celebs presumable, Courtney Love passing out backstage was such a known thing that it didn't bother to filter it out
|
|
#
?
Oct 21, 2022 05:14
|
|
|
The best way to prevent people from getting hurt is to release these models publicly so that everyone understands what AI is capable of. The most dangerious outcome is when the CIA and advertising companies are the only ones who have access to image AIs. If the AIs are secret then people will be much more likely to believe AI produced bullshit and propaganda. Making it public inoculates the public against believing every picture they see.
|
|
#
?
Oct 21, 2022 09:26
|
|
|
precision posted:presumable, Courtney Love passing out backstage was such a known thing that it didn't bother to filter it out Courtney love isn't important enough for anyone to add her to the list
|
|
#
?
Oct 21, 2022 09:27
|
|
|
BoldFace posted:Stability.ai's CIO made a blog post and answered some questions on Reddit. They maintain that 1.5 was leaked by RunwayML, which they call "a bad faith actor" who reneged on promises made privately. Someone from Stability.ai did request HuggingFace to take down the model, but Emad on discord clarified that it was not a "legal takedown request". Later they dropped the takedown request altogether and it seems that the leaked model can stay up. I guess this proves that the claim by stability AI that anybody who worked on a model can release it at any time was always bullshit. If any employee did it, they almost certainly would have been fired and possibly face legal consequences.
|
|
#
?
Oct 21, 2022 14:14
|
|
|
I like this handsome art deco elf.
|
|
#
?
Oct 21, 2022 14:23
|
|
|
I'm not seeing a huge upgrade with 1.5 in any case. Lots more steps but the results aren't dramatically different, other than those inpainting improvements. I wonder if this is about as far as this model can go.
|
|
#
?
Oct 21, 2022 14:31
|
|
|
Mozi posted:I'm not seeing a huge upgrade with 1.5 in any case. Lots more steps but the results aren't dramatically different, other than those inpainting improvements. I wonder if this is about as far as this model can go.  But if you try the leaked anime model, it's clear how far we are from approaching even its performance. Huge difference just from how much better the training images are tagged. Potential big gains also on the natural language model, DALL-E seems to be quite a bit better at it already.
|
|
#
?
Oct 21, 2022 14:51
|
|
|
|
|
#
?
Oct 21, 2022 15:04
|
|
|
"highly detailed and realistic oil painting of an Elvish Necromancer sitting at a desk. on the desk is an adorable cat"  clicking Variations on the first one gave me this absolute stunner  if nothing else, this is a great tool for nerds' D&D campaigns. suddenly you can have a picture ready to go for every encounter your players face. that's so cool. anyone wanna play D&D with me, send me a PM. i'll DM edit: i'm serious, i really wanna do this now lol
|
|
#
?
Oct 21, 2022 15:13
|
|
|
precision posted:if nothing else, this is a great tool for nerds' D&D campaigns. suddenly you can have a picture ready to go for every encounter your players face. that's so cool. anyone wanna play D&D with me, send me a PM. i'll DM https://forums.somethingawful.com/showthread.php?threadid=3654812 I've also considered making a thread in /Tradgames specifically for AI art we made for our games, but I'm too lazy to write the OP.
|
|
#
?
Oct 21, 2022 15:15
|
|
|
Some spooky space angel skeletons... Something lost. A shape remembered.  The sound of the machine.  Promethean clockwork.  Form unbounded.  Both winged guardian.  And blessed demon.  Watchers from the ramparts.  But even Angels fall.  To ascend anew. 
|
|
#
?
Oct 21, 2022 16:05
|
|
|
uh, are you okay, DALLE?  "realistic photograph of a teenage slacker being yelled at by their mother while standing in a kitchen"  
|
|
#
?
Oct 21, 2022 16:47
|
|
|
precision posted:uh, are you okay, DALLE? Good prompt lol. This is same one with SD, though it looks more like a step-mother situation 
|
|
#
?
Oct 21, 2022 17:07
|
|
|
mobby_6kl posted:It seemed to create reasonable interpretations of the prompt a bit more often too, but it's hard to quantify How much of this is better tagging versus just having narrower goals? But yeah, the NAI model is pretty impressive, if you're using it to generate at least vaguely anime styled pictures of people. You just need to be aware that you're giving instructions to an artist who is doing their best but is also very horny.
|
|
#
?
Oct 21, 2022 17:10
|
|
|
precision posted:
I spent an a hour or so generating a bunch of random NPC portraits for my rogue trader game. I can't get it to give them proper skull adorned clothing, but generic gothic sci fi nobles are good.
|
|
#
?
Oct 21, 2022 18:09
|
|
|
The big improvement in v1.5 is supposed to be more accuracy in the small details that 1.4 sucked at like janky hands/feet and maybe faces? It's been available online through dreamstudio for a few weeks already, this model release is the same thing but now you can run it locally. There are already some redditors claiming that it's been trained to be intentionally worse at drawing children to combat CP:  Of course no comparisons or evidence posted but they're still getting mad
|
|
#
?
Oct 21, 2022 18:35
|
|
|
Emad has said that they will not release NSFW models going forward. 
|
|
#
?
Oct 21, 2022 18:40
|
|
|
BoldFace posted:Emad has said that they will not release NSFW models going forward. Now that the SD source is out there, training a new model only takes time and processing power. This seems like an opportunity for another company to step in and offer an uncensored model. I don't particularly care about making porn with these things, but I don't like the idea of an arbitrarily censored model. It would be like buying a copy of Photoshop that won't let you edit naughty images.
|
|
#
?
Oct 21, 2022 19:39
|
|
|
WhiteHowler posted:Now that the SD source is out there, training a new model only takes time and processing power. This seems like an opportunity for another company to step in and offer an uncensored model. The processing power is the largest expense by far. sd v1.4 took $600,000 worth of rented cloud GPU time. Which is still a huge improvement compared to the expense Google or OpenAI went to for their own models! 1.5 is I think significantly more training time (= more expensive). All the models like waifu-diffusion and novelai etc are just fine-tunes of 1.4 and not entire new models. It would be very expensive to replicate.
|
|
#
?
Oct 21, 2022 19:47
|
|
|
Brutal Garcon posted:How much of this is better tagging versus just having narrower goals?  Could be part of the explanation, but it doesn't seem to be limited to just naked anime girls. Like that yelling prompt from above, these was among the first 6 or so I think.  And it can still do other stuff too so it's not like it's completely crippled.     Just seems like with any machine learning problem, the biggest challenge is having good data and I feel like this isn't an exception. It's going to be difficult to overcome too since you need millions of well labeled images. RPATDO_LAMD posted:The big improvement in v1.5 is supposed to be more accuracy in the small details that 1.4 sucked at like janky hands/feet and maybe faces? It's been available online through dreamstudio for a few weeks already, this model release is the same thing but now you can run it locally. BoldFace posted:Emad has said that they will not release NSFW models going forward.
|
|
#
?
Oct 21, 2022 19:48
|
|
|
RPATDO_LAMD posted:The processing power is the largest expense by far. sd v1.4 took $600,000 worth of rented cloud GPU time. A few years ago this was millions, $600k is crowd fundable and it will keep dropping.
|
|
#
?
Oct 21, 2022 19:49
|
|
|
Needed some Space Miners. AI seems to think that if they mine space...must be space in that cave. Torch.  Egg found.  Structures.  Always time for a selfie? 
|
|
#
?
Oct 21, 2022 20:17
|
|
|
i was going to do a joke "AI porn" post with the prompt "two attractive and empowered women appreciating each other" but EVEN THAT IS CONTENT BLOCKED LMAO "two attractive and empowered women smiling at each other" worked, but didn't produce anything funny
|
|
#
?
Oct 21, 2022 20:47
|
|
|
does it understand "mirin' "
|
|
#
?
Oct 21, 2022 20:48
|
|
|
Alright I got the 1.5 model close up of a bride and groom's hands at a wedding, wedding ring, exchanging vows, exchanging rings negative prompt: lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, [blurry], CAPTCHA, (((deformed))), bad anatomy, disfigured, poorly drawn face, mutation, mutated, (extra_limb), (ugly), (poorly drawn hands), messy drawing  From this tiny sample size it looks like it's better at hands but still very far from perfect.
|
|
#
?
Oct 21, 2022 20:53
|
|
|
if yall didn't see it i'm doing a CYOA with DALLE images https://forums.somethingawful.com/showthread.php?threadid=4015378 it's not a joke, as long as people want it to go on, i'm down for putting some real effort into it the prompt used to get the first four faces was just "a teenager in 1994". i've noticed that when you use vague terms, you always get an almost perfectly "diverse" response, so that it's kinda like a Gap ad. DALLE clearly, in some way, understands "diversity" and also seemingly understands that humans desire it in the output because there's no way they told it to be diverse
|
|
#
?
Oct 21, 2022 21:11
|
|
|
^^^ https://twitter.com/waxpancake/status/1549076996935675904 precision posted:i was going to do a joke "AI porn" post with the prompt "two attractive and empowered women appreciating each other" but EVEN THAT IS CONTENT BLOCKED LMAO And the anime model is... well, who's to say that two women can't be just.. very close to each other? https://i.imgur.com/2VgPpzW.png RPATDO_LAMD posted:Alright I got the 1.5 model The Novel model is generally pretty good about hands in normal images but if you try to make it do closeups, it's all absolute nightmare sausage fingers  The negative prompts do help though
|
|
#
?
Oct 21, 2022 21:14
|
|
|
precision posted:because there's no way they told it to be diverse they actually do, it detects words that are vaguely about people, then add a random race to the end of your prompt this got discovered by people noticing false positives, then putting in prompts like 'a stick person holding a sign that says:' and seeing what was on the sign.
|
|
#
?
Oct 21, 2022 21:16
|
|
|
|
| # ? May 29, 2024 15:00 |
|
|
Megazver posted:does it understand "mirin' " I'm gonna say yes, slightly two bros, gym outfits, face to face, post-workout  two bros mirin', gym outfits, face to face, post-workout  precision posted:if yall didn't see it i'm doing a CYOA with DALLE images This is a pretty cool thread concept! And yeah Dall-E manually adds diversity keywords to the prompts behind the scene, you can tell what kind of keywords it's adding with some clever prompts: https://twitter.com/rzhang88/status/1549472829304741888
|
|
#
?
Oct 21, 2022 21:20
|
|