|
Dr Christmas posted:Has anyone come up with a term for how the nod towards realism or painterly-ness or a specific art style gets AI art people to overlook stuff like the supposedly lavish details just being scribbles, or getting the dang number of limbs wrong? The Cheesecake Factory Effect
|
# ? Feb 27, 2024 07:24 |
|
|
# ? May 30, 2024 11:10 |
|
|
# ? Feb 27, 2024 08:08 |
|
That bear needs to put on her pants.
|
# ? Feb 27, 2024 08:24 |
|
Snowglobe of Doom posted:Don't forget that all this happened in Glasgow so if there was any Oompa Loompa singing it would have been in a angry Scottish accent I'm sorry, but as someone from Scotland I had to correct it.
|
# ? Feb 27, 2024 08:35 |
|
They promised an experience you've never had before. I'd say they clearly delivered. Gotta say though, hanging up giant prints of the AI ads as an exhibition piece is one ballsy move.
|
# ? Feb 27, 2024 08:42 |
|
Yeah really. Everyone was very surprised, as advertised
|
# ? Feb 27, 2024 13:48 |
|
Cloacamazing! posted:They promised an experience you've never had before. I'd say they clearly delivered. Look, they were promotional images, and those same promotional images were in fact at the show. You can't sue. In fact, we should sue you for being mad about it.
|
# ? Feb 27, 2024 13:50 |
|
Come for the cartchy tuns, stay for the sweet teats
|
# ? Feb 27, 2024 15:49 |
|
https://twitter.com/DiscussingFilm/status/1762503045940724112?t=f5BiK-PPdVEr_7CWdGduew&s=19
|
# ? Feb 28, 2024 01:22 |
|
No thanks, I prefer the natural human exertion scent I've built up over the years
|
# ? Feb 28, 2024 01:30 |
|
I've always wondered what Bloodborne would smell like
|
# ? Feb 28, 2024 01:42 |
another thing i am wondering is what do you think the orphan of Kos smells like haha im just curious for laughs haha
|
|
# ? Feb 28, 2024 02:03 |
|
I think mine is broken, it keeps releasing the napalm smell whenever I beat a level?
|
# ? Feb 28, 2024 03:29 |
|
|
# ? Feb 28, 2024 04:04 |
|
AI is too worthless to do anything of value but just functional enough to destroy any remaining semblance of trust or credibility for anything across the entire internet cool love that for it
|
# ? Feb 28, 2024 04:08 |
|
Searched Google Images, the first row of results are all real but this was in the second row along with the pictures in that tweet. https://medium.com/@scottishfoldmunchkin/can-a-serval-cat-kill-a-human-ef282d94903c Whole article and accompanying pictures are AI nonsense, the other ones from the same author are even worse.
|
# ? Feb 28, 2024 04:23 |
|
Here's the blog site that generated those fake tufty Cerval images: https://conservationcubclub.com/ It seemed to start out as a regular blog about cats with actual photos but recently they've bee using AI images pretty much exclusively and for some reason they like to add ear tufts to most of the cats. They're usually otherwise passable but some AI weirdness still manages to creep in here and there:
|
# ? Feb 28, 2024 04:42 |
|
NoEyedSquareGuy posted:Whole article and accompanying pictures are AI nonsense, the other ones from the same author are even worse. LOL that car
|
# ? Feb 28, 2024 04:45 |
|
Snowglobe of Doom posted:Here's the blog site that generated those fake tufty Cerval images: https://conservationcubclub.com/ Oh hi Josie!
|
# ? Feb 28, 2024 05:03 |
|
ScienceSeagull posted:Oh hi Josie! Lol, drat Also hooray for the internet's continued march into being negatively useful, I guess
|
# ? Feb 28, 2024 05:05 |
|
The Sausages posted:Most of the elaborate prompt-fu is cargo culting - people are copying phrases and words that more often than not don't affect the output, because they saw it in someone elses prompting and it seems logical to put stuff like "masterpiece" as a prompt or "bad anatomy" in the negative. But the models/loras/embeddings used affect the style and composition far more than "masterpiece, bad anatomy" do, and some things that seem sensible to AI creators are provably meaningless to the AI e.g, iirc "extra limbs" has zero effect on the number of limbs when included in a SD1.5 prompt, no matter how many (()) you include, Unfortunately you need a lot of pictures and a lot of descriptions. Because SD and midjourney etc are plagiarism machines they did this by stealing almost every image on the Internet and using text scraped from the pages they found them in as data to automatically generate the tokenised descriptions. Typing "masterpiece" will bias toward results plagiarised from better artists because people wrote "A masterpiece!!!" as comments under them more often than other images, as well as stuff like competitions and image tag sites having those as specific categories. Negatively weighting "bad anatomy" does actually work in plagiarism-heavy models because people repost bad art and call it out. Similarly scraped medical images means "Polydactylism" can reduce finger problems. None of these guys know this, they're just copy-pasting stuff from the first Google result of "Stable Diffusion negative prompts porn best", and yeah exactly what works as a magic token varies wildly between models, releases, and whatever those specialised addon models are called I forget. It's pure cargo culting but at some point the planes were real. Disclaimer: The technical parts of this post have been compromised for readability. Also the technical methods used and the provenance and quality of the source data are far from the only reason why pattern matching as a source of generating imagery without understanding the fundamentals is a flawed concept. Splicer has a new favorite as of 13:17 on Feb 28, 2024 |
# ? Feb 28, 2024 12:38 |
|
Who the hell looked at the internet of all places and went "there's just not enough cat pics, we'd better generate some with ai..."
|
# ? Feb 28, 2024 12:46 |
|
Snowglobe of Doom posted:Okay it turns out that I owe AI apps an apology because whatever app this moron used wasn't actually responsible for the terrible anatomy and pose and facial expression on that model since it turns out the dumbass drew all that terrible amateurish poo poo himself and used AI to add details:
|
# ? Feb 28, 2024 13:25 |
|
Stupid_Sexy_Flander posted:Who the hell looked at the internet of all places and went "there's just not enough cat pics, we'd better generate some with ai..."
|
# ? Feb 28, 2024 13:28 |
|
The Saddest Rhino posted:https://twitter.com/DiscussingFilm/status/1762503045940724112?t=f5BiK-PPdVEr_7CWdGduew&s=19 Finally we will know what Yoshi’s eggs smell like
|
# ? Feb 28, 2024 14:30 |
|
There are so many facebook groups that just post endless bullshit AI pictures for Boomers to comment "WOW!!!" on. It's bizarre. Even the really obviously fake stuff has loads of comments that may or may not be bots posting on them how much they want whatever it is. The internet seems to have got so so much worse over the last year or two.
|
# ? Feb 28, 2024 14:37 |
|
Apologies for nitpicking but that's how GAN image generators work. Diffusion Models such as SD, MJ or Dall-e are trained by adding noise to an image in a way that allows the AI model to learn to reverse the process. "Plagiarism Machines" is an excellent and accurate description in both cases. The other nitpick is the text used for the SD dataset was the alt text i.e. what gets displayed if the image doesn't load, and not the text from the page or comments. Apparently you can search the dataset here to get an idea of the GI part of AI's GIGO, or to see if your pics are included in the dataset. There's also another anime/hentai trained dataset ripped from danbooru or something that had more accurate captioning and some of that has spilled into people's prompts e.g. 1girl, or pretty much anything with underscores. So "extra_limbs" might have an effect as a prompt (positive or negative) with the Waifu models, but both "extra_limbs" and "extra limbs" won't be effective with SD models. For Dall-e 3 the dataset was re-captioned by AI image analyzers; MJ and SD are probably following suit as fast as their budgets allow. Harvey Mantaco posted:AI is too worthless to do anything of value but just functional enough to destroy any remaining semblance of trust or credibility for anything across the entire internet cool love that for it
|
# ? Feb 28, 2024 14:47 |
|
Someone posted a picture of what was supposedly Garden of the Gods and my mom shared it with my brother and I. I wasn't sure if it was AI or just some kind of Photoshop composite image but it had some auroras in the background but it was also apparently sunset and the rocks had light shining on them from a completely different horizon than where any of the other light in the photo was. My brother and I both saw it immediately and were like "that's obviously fake" and she got upset because she thought it was beautiful and we were being rude by pointing it out I feel like a lot of older people online think of things this way, where the sense of something is completely secondary to whether or not it's kind of surface-level pretty, like a Kinkade piece where it's just a vaguely nostalgic mishmash of busy color that resembles a landscape if you squint
|
# ? Feb 28, 2024 14:47 |
|
The Sausages posted:but that's how GAN image generators work. Diffusion Models such as SD, MJ or Dall-e are trained by adding noise to an image in a way that allows the AI model to learn to reverse the process. The second I was just wrong. I also overestimated what was involved in tokenisation - I thought there was synonym work put in but apparently not. I'd like to claim I was merely underestimating their laziness but nah, I just took stuff I heard as fact without checking. The rest was stuff I deliberately left out because of length but well worth posting! The Sausages posted:Apologies for nitpicking
|
# ? Feb 28, 2024 15:34 |
|
i ran into this myself and it sucks so tremendously badly. i was working on an art piece and needed reference pictures of crocodiles (skin, teeth, etc.) and a good half of the results on google images were ai genrated pure garbage that's worse than useless for reference who even looks at these things and think "yeah this is good, i'll post this on-line"
|
# ? Feb 28, 2024 16:22 |
|
Your Computer posted:i ran into this myself and it sucks so tremendously badly. i was working on an art piece and needed reference pictures of crocodiles (skin, teeth, etc.) and a good half of the results on google images were ai genrated pure garbage that's worse than useless for reference Terrible crocodile, not bad Crystal Palace iguanodon
|
# ? Feb 28, 2024 16:28 |
|
Your Computer posted:who even looks at these things and think "yeah this is good, i'll post this on-line"
|
# ? Feb 28, 2024 16:56 |
|
|
# ? Feb 28, 2024 16:58 |
|
I want a shirt that is just the short sleeves, nothing else
|
# ? Feb 28, 2024 17:04 |
|
My favorite restaurant Modlidani's
|
# ? Feb 28, 2024 17:22 |
|
Your Computer posted:i ran into this myself and it sucks so tremendously badly. i was working on an art piece and needed reference pictures of crocodiles (skin, teeth, etc.) and a good half of the results on google images were ai genrated pure garbage that's worse than useless for reference very cool that later generations of AI will be trained on the output of current gen like this
|
# ? Feb 28, 2024 17:28 |
|
Large Cokes wore jean vests and smoked indoors?
|
# ? Feb 28, 2024 18:39 |
|
My hometown had the last Modlidani in the country, next to the K-Mrrkt. Always wondered by they served their drinks in fry cups but nowadays I miss that simplicity. McDonald's won't do if even if you ask.
|
# ? Feb 28, 2024 19:09 |
|
Heath posted:Someone posted a picture of what was supposedly Garden of the Gods and my mom shared it with my brother and I. I wasn't sure if it was AI or just some kind of Photoshop composite image but it had some auroras in the background but it was also apparently sunset and the rocks had light shining on them from a completely different horizon than where any of the other light in the photo was. My brother and I both saw it immediately and were like "that's obviously fake" and she got upset because she thought it was beautiful and we were being rude by pointing it out Can you repost it? I live near GotG so I'd love to see it.
|
# ? Feb 28, 2024 19:55 |
|
|
# ? May 30, 2024 11:10 |
|
It's way back in my group chat but I'll try to find it
|
# ? Feb 28, 2024 23:30 |