#ai weirdness

[alovelessmelancholy]

Double the fun🔥

[maxwellshimbo]

Botober, day 1: Bread sky

[benthesoldiersjeanshorts]

[seriously-mike]

Extremely Stupid AI-Generated Shit

(that is still kinda funny, anyway)

Those little freaks are the result of the following prompt:

Glurb snorf thwip krazl vomp yurgle zibble frunx quorl plimf drax gnurk jibbit flox zark welp thrum skork plund frazzle mreep

Top image comes from Midjourney, the bottom two are probably DALL-E 3 (the last is certainly DALL-E 3, the middle I'm not sure but it does look like it). To make this even weirder (and funnier), Bing Image Creator considers "Glurb" an unacceptable word.

Okay. I did refer to oblong, roundish, organic shapes as "blorps" a couple of times, but this looks like someone posted his kid's drawings of weird critters on the internet a long time ago, the algorithms yoinked them unceremoniously along with the descriptions, and just like that red t-shirt that turned the entire load of your laundry pink that one time, weird kid drawings pounded into mathemagical fairy dust along with more typical fairy tale and fantasy illustrations resulted in the weird names assigned to... this.

This is merely a selection of pics generated from this prompt, but the overall concept tied to it are creepy round-bodied creatures for Midjourney, goofy cartoonish Monsters Inc. for DALL-E 3, and...

I just scrolled through the post and found results for various Stable Diffusion data models. And Stable Diffusion, ladies and gentlemen, consistently responds with goblins.

This Warhammer miniature-styled thug fell out of Stable Cascade, the weird semi-forgotten uncooperative child. For the result of a string of completely nonsensical words, he's surprisingly coherent, with a fairly regular number of fingers AND toes. Of course the details like his kneepads are still blorpy, but that's how Stable Diffusion rolls, even three years, four major versions and a shitton of fine-tuned custom models in.

And SDXL custom model called FenrisXL provides an entire fucking family of goblins. What is going on here, because my assumptions regarding Stable Diffusion and SDXL in particular just have been challenged.

First, the Kitten Effect is less pronounced than it was in the early versions of the algorithm, if it happens at all. I'll chalk it up to improvements in the XL algorithm. Second, they're cartoonish goblins, but the Same Face Syndrome usual for the XL algorithm (every fucking custom model I tried suffers from it, no ifs, no buts) is less pronounced here than it is in case of human characters. Third, how in the FUCK an entire family of goblins spewed forth from a prompt consisting of gibberish has almost perfect and repeated anatomy, not counting the orphaned hand on the goblin girl's shoulder and an extra toe on the guy second from left in the front row? And varied skin and hair colors?

I can only explain it with someone lucking out on the seed number, much like I lucked out on the entire Chinese Garden test last year.

Still, though. Goblins. Fairly solid in custom models, messier in the core SDXL 1.0 (below), without any meaningful words in the prompt.

Where the fuck are they coming from? This is some serious Horse K shit and I refuse to investigate it any further. Much less add other weird phrases like "Yakka foob mog!" or "Kov schmoz, ka-pop?" to it and test it on my build (or even Photobooth from Hell in particular). It's late and my brain is giving up.

[topoillogical]

Lately I've taken to torturing AI copies of karkat by putting them in a room together

And they eventually worked themselves all the way to the cancer metaphor (I'm so proud....)

[pixelsynthesis]

[terracebatman]

Asking AI who would be a couple in the afterlife? Mr Rogers & Princess Diana.

[quihi]

Sometimes I think the surest sign that we're not living in a simulation is that if we were, some organism would have learned to exploit its glitches.

Janelle Shane, You Look Like a Thing and I Love You

[maxwellshimbo]

Botober, day 3: Exceptionally mischievous pumpkin

[benthesoldiersjeanshorts]

[beepboopappreciation]

Is this anything

[brucesterling]

*Is it true, or did Google Gemini just make it up, or did somebody just make that up about Google Gemini?

*Well, who knows; your guess is as good as Google Gemini's

PS: How many of those apparent Gemini answers are faked, made up by humans out of spite, or for the lulz? Plenty.

*And there will be plenty more.

*This just in: Google AI lamenting that, although their machinery does spit up some very weird stuff under weird circumstances, mostly they're being libelled by forged memes.

[seriously-mike]

I don't want to state the obvious, but AI is dumb as shit.

I wanted to try and generate a photo of an interesting Old West gentleman, so I tried a prompt like this:

vintage portrait photo of a gaunt old man with shoulder length white hair and mustache, wearing a black suit, standing in the desert, lee van cleef, [[christopher lee]], [peter cushing], lance henriksen

(square brackets mean less emphasis on those two)

I got the photo on the left as the result. Looks perfectly generic, right? But, just to be sure, I added "Sam Elliott" to the negative prompt and...

I got the guy on the right. Three times out of four in one batch. How in the fuck does the algorithm combine "old", "mustache" and "desert" into one "Sam Elliott" and remove those elements from the photo altogether? This is some Švejkian shit going on: "you asked about an old mustached guy in the desert who isn't The Old Mustached Guy In The Desert, Sir. I know no other one, the best I could find is this young fellow."

For the love of God, I'm asking about four different guys that can be averaged into one character and should have their photos all over the fucking internet, ready to be stolen. The AI insists on giving me a fifth guy without asking and gets very upset when I catch it doing it.

[seriously-mike]

Yes. It's a novelty trick but it quickly gets boring.

Okay, I still get a kick from listening to my lyrics wrapped in some ridiculously pop AI-generated music, and it still gets views, and there's ten bucks on the table that would get those tunes on Spotify because people actually want it, but it's a fad. Plus, the tech is rickety as fuck and I'm putting up with it only because I know how much of an expensive pain in the ass being in a band is, and I know my vocal capabilities (not great, absolutely terrible - and there's video evidence for it).

And for AI-generated images, the hype is dead or at least effectively dying. Just like I predicted half a year ago. Not because some miraculous program that stopped the evil corporations from stealing everyone's art (I would like to remind you again, that one doesn't work for shit), but because these days, only the boomers, the scammers and the pervs are using it. No matter how many images they run on, no matter how large the dictionaries, image generators are softly clonking against a glass ceiling, unable to consistently draw hands properly after three years, four major versions and countless finetuned custom models cooked by basement-dwelling nerds. AI-generated videos? Forget about it. The final nail to that coffin was a has-been rock band paying someone actual money for a laughably bad montage of jittery slow pans across StableDiffusion-generated shots that kept coming apart. If you want to see that train wreck, I'm talking about Saxon's video for their new song "Madame Guillotine".

But most importantly, the main function of LLMs isn't a "sometimes-wrong talking encyclopedia". It's composing grammatically correct sentences and that is it. Trying to make it act as an encyclopedia or any sort of factual reference is NOT what it was made for. Remember thatcan.be? Or WhatWouldISay? Those things were the first attempts at what we're seeing now in ChatGPT and other LLM-based chatbots. Hell, there's a whole-ass Tumblr blog dedicated to AI weirdness (unsurprisingly, it's called @aiweirdness), and there's a whole lot of shit there that is more likely to make you question your sanity than to encourage anyone to invest in AI.

Related

[maxwellshimbo]

Botober, day 5: Skeleton with TOO many bones