AI Image Generation Algorithms - Breaking The Rules, Gently
Вставка
- Опубліковано 29 вер 2024
- One of the pieces of common advice for using image generation machine learning algorithms is 'don't ask for text output'. Well, that seems to me more like an invitation to try...
Check out Simon Roper's channel: / @simonroper9218 @simonroper9218
Join the Atomic Shrimp official Discord server for early access to videos! - / discord
Atomic Shrimp subreddit: / atomicshrimp
I think we'll look back fondly on this early era of ai art generation, with all its weird faults and quirks.
You mean a veritable wasteland of surrealist horror?
So are these strange AI generated images basically what you see when you close your eyes?
TBH I wouldn't be surprised if I saw an AI generated eldritch monstrosity in one of your videos
I read this like Salad Fingers was saying it in my head
Thanks, Bingus.
Bingus the boosing wockis is one of my favourite fairytale characters.
Would be a great album cover
Someone needs to write the book
@@molybdomancer195 chatGPT could write the book
My grandfather used to read it to me when I was small.
@@molybdomancer195 I'm on it 👍
I, too, am subscribed to Simon's channel. I also thought the generated signs seemed like olde English. So cool you got Simon to read them.
So delighted to see the shout out to Simon roper. He is one of my favourite UA-camrs. His brain is a brilliant place.
A crossover with Simon and Shrimp, excellent!
I'm glad to be reminded of Simon roper. Great channel I've not watched in a while, really interesting.
I think what impresses me the most is your ability to pronounce these fake ai words!
To me, that's fuel for nonsense words when it comes to scambaiting.
The last bit by Simon sounded like the Words of Unmaking from the movie Excalibur. Fitting I suppose.
Brilliant! Came here via a link in the comments section to a Simon Roper vid and glad I did.
Here is a poem Bing AI wrote about Bingus the Boosing Wockis:
Bingus was a boosing wockis Who loved to play and frolic He had a fluffy purple fur And a long and curly tail He lived in a forest of candy trees Where he ate the sweetest fruits He made friends with all the animals And sang them happy tunes But one day he met a hunter Who wanted to catch him for his fur He chased him through the forest With a net and a gun Bingus ran as fast as he could But he tripped on a root and fell The hunter was about to grab him When he heard a loud and angry yell It was Bingus' friends, the animals They came to rescue him from harm They scared the hunter away with their roars and bites And hugged Bingus with their arms Bingus thanked them for their help And said he was glad they were his friends They all went back to the candy trees And lived happily ever after. The end.
Entertaining as ever. It's amazing how far these have progressed in such a short time.
I have just recently seen The Grand Budapest Hotel and now get Boy With Apple! Thank you for your references.
I've had dreams that looked like that, you could identify text that really wasn't.
But also had dreams were I was able to read text very clearly.
I love the emergency instructions one, it feels like that one “how is prangent formed” video
I want to learn more about Bingus the boosing Wockus. Seems like a nice chap
you did not have to read those words so professionally, but you did, and that's amazing
5:46 "SAD JARLE" this made me laugh so much XD
1:30 which version of Stable Diffusion? That is very important question. And different samplers too can create different results
This reminds me of those sentences where they jumble every letter after the first one like it looks right at a glance but turns into gibberish when you actually try to read it as is.
man... sometimes those fortune cookies speak facts. "I wooee forite us it not thoigtt the the it wole tute" really connects with me, i feel that it really speaks to me on a personal level.
The AI generated texts gave me flashbacks to the red room scenes in Twin Peaks.
The video someone made where they fed every line of Bo Burnham’s “Welcome To The Internet” into Midjourney displays this really interestingly. I have a top comment on that vid with timecodes for all the not-quite-text
Similar neural networks with larger numbers of neurons can generate text in images fairly well, so the archetypal parts of language might be a thing.
I feel like generating poems and such using AI like the cheese poem would be a great method of coming up with a lexicon for a conlang.
I completely lost it at bingus the boozing wockis
Sounds like a Douglas Adams character
2:04 lets hope these mouse codes work better than the previous ones.
I actually love doing prompts such as "album cover: 'Lovely Cucumbers by Iron Maiden'". I also do it with movie posters/DVD box art, and video game cover art. Mixing something nonsensical with an already established IP is often pretty funny.
The cheese poems immediately made me think of Vinesauce trying to speak corruptions.
"Bingus: The Boosing Wockis
Truly a masterpiece novel of our generation
1:00 Bullsquid
5:10 Vogon poetry
Doing unusual things with things is my favourite thing to do 😁
I always liken these things to an octopus's understanding of the patterns it itself is making to camouflage itself against a background: there is no understanding. It is just camouflage.
A good example of this is looking at the flowers from 2:23 to 2:55 or so. If you're looking at the sunlight and the glass and the refraction, the camouflage is good, but the flowers themselves are, in every instance, wrong. Most of them even include weird refraction or extensions of glass and water in the "flowers".
I think if you were to look at those pictures from the perspective of being a flower yourself, they look every bit like the body horror that you get from "pencil drawing of an athlete" that you got with earlier image generators.
Another very, very important (you know, IMHO) aspect of these AI generated images is that for now they're trained on *real* images. As the internet becomes progressively more populated with AI generated images, the training set (i.e. the internet scraped for images) will include the cumulative error of previous generations of AI images. It will be like layering octopuses on top of octopuses stacked down to what you wanted them to imitate. What would that look like? Dunno.
-edit (TL;DR portion)
In particular, lets look at the glass lobster picture at 3:00 just a bit closer. You might notice a few things wrong:
1, the lobster has 3 antennae that merge and branch in an unlobsterlike way.
2, the left side of the pine table (with the lobster as a divider) is a different color and grain than the right.
3, the grooves in the table don't line up.
4, the shadow of the lobster is wrong. Its not in sunlight. It looks like a diffuse shadow that's an artifact of most glass sculpture images being against a white background with diffuse light. It looks like a bad green screen shadow.
5, what's that little green thing in the window? Well, its a nonsense. There's supposed to be "stuff" visible through windows and it put "stuff" there.
At first glance it does look convincing in exactly the same way that you would never notice a cuttlefish nestled in against some coral and sand or a leaf butterfly on a stick at first glance, but even a mildly thorough look reveals any of those things to be not what they seem.
One of the reasons these things excel at artistic impressions over photorealism is because all of those things that I mentioned could just be chalked up to "artistic expression", though I suspect that if I were an accomplished oil painter, those images would also fail in obvious ways not evident to me, me being inexperienced with oil painting.
I'd really suggest taking a closer look at the things in the background, the continuity of objects, whether windows on the same building are the same style. Not a single one of these "photorealistic" images hold up if you're given more than a couple of seconds to look at them.
The short poem about cheese just makes me think that this would be a brilliant way to create made up languages for fantasy settings
Gosh so we have Simon Roper referenced, and David Firth in the comments. This is my corner of the internet.
Also "ARRSK" is a smooshing of ARGH and RISK and ARSE and may very well be the sort of thing I'd exclaim in a moment of panic.
Great video, randomness i didnt know i needed
Simon's readings were awe inspiring, yours however were simply hilarious.
As someone who isn't an artist, I welcome the new AI art revolution. Theyll get over it!
You are your videos make me so hopeful about life.
I lost it at Bingus the Boosing Wockis
the images are likely to be from all countries in the world and are therefore all in different languages
I’d imagine having a stroke and losing the part of your brain that can read and write would be like looking at these images of text. They’re both familiar yet alien and unintelligible at the same time; strange indeed.
Those mice and squirrels look creepy.
The first chapter of Bingus the boosing wockis needs to be written.
I need that bingus the boozing wockus image in 4k
The cheese poems sounded like they were being spoken in a foreign language
This was great fun.
I interpret tongues at church. So the last one says "you are the son I imagined I would get if I got things right"
I wonder what it would generate if asked for an oil painting of Mr barrister John warosa.
I think we will look back on this era as "Rubbish in, Rubbish out", when "AI" is eventually recognised as unattainable. Very funny! (I love Simon too, both of you are what UA-cam was created for).
"DO'T AST DX'T MAS TET"
I'd love the world where the ai's are unlimited unrestricted and free, only with ads.
It’s like text when you dream and try to read itv
Dear ChatGPT, please generate a poem, in old English, that Atomic Shrimp might ask of Dall.E or Stable Fusion, in the performance style of Simon Roper.
Given how many misspellings of “yard sale” already exist, I wouldn’t be surprised if “YARLD SALE!” was a real one too
Could try new scrambled versions of the FAULTY TOWERS sign.
YARD YAY
Yes! Made me think of a well-known example "YARD SARD"
YARRD OR SALD
those were absolutely the most true-to-life images ive seen from any neural network
I genuinely laughed like an idiot at 'Bingus the Boozing Wockis'. That definitely needs to be a thing.
Needs to be featured in one of those precautionary advertisements that Shrimp does. :)
I'm laughing my ass off on my way to work replaying that clip I can't breathe
I'm crying in laughter sitting in a café looking at that creation
Maybe he can ask chatGPT to write this story in a follow up video :D
@@Soken50 I prompted chatgpt to write a story about him.
It goes something like this:
"Once upon a time, in a far-off forest, there lived a Wockis named Bingus. Bingus was not like any other Wockis in the forest, for he had a love for one thing above all else: booze.
While other Wockis spent their days foraging for food and building homes, Bingus spent his time searching for anything he could ferment and turn into alcohol. He collected berries, fruits, and even flowers to make his potent concoctions.
Bingus was not a very good Wockis. He often stumbled around drunk, and his home was always in disarray. His fellow Wockis disapproved of his behavior and shunned him. But Bingus didn't care. He was content with his boozy lifestyle.
One day, while out searching for ingredients for his latest brew, Bingus stumbled upon a group of humans having a party in the forest. The humans were drinking and dancing, and Bingus couldn't resist joining in.
As the night wore on, Bingus became the life of the party. His drunken antics and strange Wockis behavior had the humans in stitches. They even started calling him "Boozing Bingus," which only made him drink more.
But as the night drew to a close, Bingus realized that he was lost in the forest, far from his home. He stumbled around, trying to find his way back but to no avail. As the sun rose in the sky, Bingus collapsed from exhaustion and too much alcohol.
Hours later, Bingus woke up to find himself in a strange place. He was in a human hospital, hooked up to all sorts of machines. The humans had found him and brought him here, hoping to save his life.
Bingus was confused and scared, but he soon realized that he needed to change his ways. He couldn't keep living his life as a drunkard, ignoring his responsibilities as a Wockis. So he sobered up, apologized to his fellow Wockis, and started to make amends for his past behavior.
Bingus still enjoyed a drink now and then, but he never let it consume him again. He learned that there was more to life than just getting drunk and that being a responsible member of his community was more important than anything else. And so, Bingus the Boozing Wockis became Bingus the Responsible Wockis, a role model for all the other Wockis in the forest."
It even has a moral!
It's always entertaining to listen to you read out gibberish very literally and with a straight face, whether it's written by a scammer or a machine.
he is experienced in reading identically nonsensical gibberish from scam/phirshing emails
I imagine there must be so many outtakes for him to do this so fluidly! And I agree, it's very entertaining!
I like it when every so often, once in a blue moon, he can’t help but laugh. I remember ‘disguised as gold bricks’ but that may be the only time.
@@ollllj hello there fellow rainbow Megute
I appreciate that right after you explained how the meanings of sentences can be misinterpreted or ambiguous, you then showed a range of photos of chicken stock as examples of "stock photos".
I had a lucid dream in which I “woke up” (while still dreaming) in a little gift shop mall. It occurred to me to look at the words written on the packages of items on the shelves, and they were much the same as seen here. Their form and presentation looked appropriate, and from a distance everything seemed fine, but up close each word was believably structured gibberish.
It makes me wonder about any parallels between the processes the mind uses to structure dreams and that used by a.i.
I know when I'm dreaming once I try to use my phone in the dream. It looks normal but clicking on apps or trying to take a picture makes everything jumble
There is probably some sort of correlation, your mind is just piecing together things it's seen and makes something believable of it, but since there's no physical relation to the imagination the things you see don't have to follow the rules of nature.
I once dreamt I was walking in the middle of a city I know very well, the difference was, it was all empty, not a single person, car or tram, and the sun was shining a little too bright, but it was a lovely stroll and everything else appeared truthful to reality.
This is a common reality check in lucid dreaming. Looking at a clock or any form of text to see if it's malformed or odd is an indication you may be dreaming, you can also try looking away and try willing the text/numbers to change and then looking back. I remember the clock check but you are right about the text too!
Yeeeeaaaah, I also always thought about that. Also while focusing on what exactly my mind imagines and comparing it to AI work
IMO the part of your brain that organizes your dreams probably can't understand language, but knows how to make things look believable. This is likely whats going on here with the AI
The “sard” sign made me laugh so hard. One of the places I frequent has yard sale fundraisers, but before the first one the person making the sign for it was half asleep. So she accidentally write “yard sard” instead. We’ve been calling em yard sards ever since.
Reminds me of my friend writing “fun facts about funland” for his geography class
I love Funland! 😃
Can we just take a moment to appreciate flawless pronunciation of Welsh double l with no pause or hesitation?
Lots of English English-speakers can do this - we've heard plenty of people pronounce Llangollen and Llandudno properly over the years! I don't doubt that many cannot (or do not, in any event). In the event that anyone does *not* know how Ll is pronounced, I give you the exposition by Edward Lhuyd (1660-1709):
L has two pronunciations in the oldeſt MSS. That which is common in the Engliſh and all other Languages, and that which is at preſent peculiar to the Welſh, and is beſt learn'd by the Ear. Its perform'd by hiſſing the L thro' the Teeth ſide-wards, and altering it in the ſame manner as T is alter'd by the addition of the h, in the word Think, &c.
@@jonathanrichards593 Brilliant!
And that X!
IKR it's llovely
in fact, without hesitation, deviation or repetition...
Made me realize how much I enjoyed the AIs just being artsy and weird in the beginning.
Being efficient and doing what they're told is probably a good thing, but I do hope we'll always have a few 'special' wacky AIs around for hilariously unintentional outputs.
I wonder if it would be possible to ask these AIs if they could create an image of ‘badly made AI generated image of a dog’ or ‘a dog in the style of AI generated images from the early 2020s’ or something akin to this. I also wonder if it would be able to create these images, and if we would have to feed it these horribly made AI images so it can learn what it would even mean to prompt these things. Then we would perhaps be able to make it create bad/wacky AI images intentionally and on purpose.
@@SophiiLuca brilliant 🤣
@@SophiiLucainfinite ai loop
They'll still be around. Many of these AI models can be run locally on people's PCs, so people will have them downloaded which means they will likely be reuploaded by people who collect all the versions as they come out.
The Ai is not there to make art,
is just to imitate the work of people so you can pay the corporations instead of the artists
so if a picture looks weird and interesting,
the developers will put a lot of effort to fix that, so it can looks just like a human work
5:56 I don’t know why but this really made me laugh. The deadpan “WANGED” just got me
Pergency Wanged, what's more. It's like something a small boy would say...
@@jonathanrichards593 that reminded me so much of that meme where no one can spell pregnancy
6:42 BINGUS????? AI generated Bingus???!
These ai-generated texts look like what I imagine text in dreams to look like...resembling real text, but without any rhyme or reason or actual meaning. Sometimes stunted or distorted. Fascinating 👍
I always hear that but I distinctly remember a dream I had a few months back where I was reading something and I could read it perfectly fine. IIRC it was some kind of horror story.
@@MysteriumArcanum I have also had one dream where I actually saw legitimate text in first person
@@sarah12232 so you could read it and it actually made coherent sense?
@@MysteriumArcanum Same, but instead of some book or story, it was a label with a name and description of a species of snail, on the tank which said interesting snail was in.
It's what text in my dreams look like, although real words do turn up as well. When I look back again, the text has changed.
Bingus the Boozing Wockis sounds like something I would like to read, actually.
Maybe an Ai could write it for us.
@Harry Satchel always drinking and getting drunk, hence the "Boozing", lmao
We need a high quality bound illustraited version of Bingus the Boosing Wockus.
Here's a possible explanation behind those patterns you saw: An image generation algorithm can't distinguish real words from gibberish, but it knows what plenty of real words LOOK like. If it sees a letter sequence like "in" or "an" or "er" over and over again, it notices those patterns and becomes more likely to generate those letters next to each other. It's also trained on lots of text with an average word length of 4 or 5 letters, so it has a pretty good understanding of how many letters usually appear before a space.
For the warning signs, it probably has a lot of training data that matches that description, with words like "danger", "caution", "emergency", "warning" etc. It knows the approximate appearance of words that are most common at the top of those signs, but not what distinguishes one word from another, hence why you got a lot of close-but-not-quite outputs like "emerercenticy" and "wanged".
The font is also significant. Warning signs are usually in all-caps, sans-serif font, which are pretty consistent and easily reproducible. A poem or proverb could be in all sorts of fancy decorative fonts, making letter patterns harder for an algorithm to detect.
Lol it’s purposely making mistakes bro. It can read text. Write an entire essay word perfect within seconds. Create detailed images within seconds. It’s safeguarding against fraud etc, i could be saying create a realistic image of showing screenshot me transferring money to u from my bank. It’s their version of having a watermark so u can see an image isn’t real.
@@YougottadoWell, I made that comment over a year ago, and AI has changed a lot since then, so my comment above was limited by what I knew about existing technologies at the time.
To your point, that's certainly one of the possible reasons - similar to how early versions of DALL-E wouldn't reliably generate faces - but it's far from the only reason.
ChatGPT may be able to "write" an essay in seconds, but from a generative perspective, text and images are two very different beasts. There are only 26 letters in the English alphabet and there are a finite number of ways to arrange them, the most common of which will be reflected in the training data. Image generation is a lot more complex because there are a lot more colours than there are letters, and there are a lot more pixels in an image than there are letters in an essay.
Ultimately, AI doesn't "know" when it produces bogus output. There are safeguards a programmer can put in place to reduce the amount of bogus output, but an AI left to its own devices is bound to mess up once in a while. Some models may be designed to produce bad text on purpose, but the longer this tech develops, the less you'll start to see that happen. A lot of generative AI models these days are open source, meaning people can exploit them for all kinds of nefarious purposes. I'd love to live in a world where it was impossible to generate a fake bank screenshot with AI, but unfortunately, that's not the world we live in.
@@YougottadoThe text AIs are different models, image AIs are just bad at text
Some people have discovered that if you ask dall-e for a written word, and then feed the nonsense word it generates back into dall-e as a prompt, you sometimes get pictures of the object the original word describes. For example, I saw someone do it with "bird". It's almost like dalle does have its own association of meaning with the nonsense words it generates, which is kind of interesting if you ask me. If you have any success with that method it could be an amusing video perhaps.
Simon Roper's reading of the poems sound a lot like the spoken dialogue in the Gravity Rush games. The spoken language in those games is a fake combination of french and japanese, and while his reading of the poems don't sound like either of those languages, it sounds so close to words you feel like you should be able to understand that it invokes the same feeling. Might just be me though, and I'll always look for an opportunity to mention my favorite games
He's pronouncing it as if it is written in Old English. This was the form of English that was used in Anglo-Saxon times... around 500AD. As an example, the Anglo-Saxon poem Beowulf was written in Old English. The scholars have worked out how the language was pronounced through studying the literary works from that period... such as by studying the rhymes in poems. It's a fascinating version of English - it's the original precursor of the language, before all the Viking and French influences kicked in, which changed the language so much. I'm sure these old languages would influence game developers, who would want to get an authentic feel for the game. :)
I was thinking Irish or Gaelic and perhaps a tiny bit of Latin and Icelandic, comparison-wise. Well or Old English of course, but that's just the boring truth.
@@happybunny1329 Love Simon's channel and also RobWords who has a video: Anglish that you might enjoy 🤠
@@LimeyRedneck Thanks very much. I will check it out. :)
@@LimeyRedneck History with Hilbert is another.
This thing creates exactly the text of messages in my dreams. Hilarious!
I think you might be onto something here. Maybe dreams are like AI, trying to make sense of the training output during the day but lacking frontal lobe sensemaking and coherence?
@@Cojrak yes. And also, where do you get your stuff? I need it.
@@Cojrak A reason text in dreams doesn't make much sense, other than maybe couple of words, is because your brains hippocampus or "RAM" isn't enough to keep it a whole page of written text in memory, and even if you have texts perfectly memorized your brain doesn't keep that text stored like an "image" that it can just load. This is true even if you vivid dream and can make perfect sense of things, when you've trained yourself to vivid dream you'll notice things can warp and change as your brain is making some approximations. So it wouldn't be related to not being able to make sense of things.
If you're looking at a map, the details will change as you move your "eyes" across the map and back, text contains a lot of very specific shapes to make sense to us and can't just be approximated, so things like text on the map won't make sense and change outside of the word you're reading. And even that might be a bit fuzzy and imperfect from person to person.
@@riggles That's actually a fascinating take. I have some education in neuropsychology but it's not something I specialize in. However, I recently read some scientific articles about inner speech, or more accurately, how very hard it is for scientist to find consensus definition and explanation for that phenomenon (there has been some research concerning schizophrenia and attempts to define voice hallucinations as inner speech of whose origins are misinterpreted, but that would require a definition of workings of inner speech so..) . So I can see your point, that text is not stored as a text but as a more salient memory, but if not text and if not speech then what is it stored as. This is more of a rhetorical/philosophical question, not directed at you, though of course if you have thoughts I would be interested in them.
I'm still left thinking about AI training input as a model/metaphor for human dreams (humans having much less and still too much input during the day). I've not thought it through scientifically yet, though I would not say it's not million miles away from Antti Revonsuo's threat simulation theory. And come to think of it, maybe it would work as a metaphor for learning that is associated in the development of psychopathology, where some events and threats are blown out of proportion due to higher than usual exposure to negative situations.
@Riggles1 I've never gotten this "text can't be read in dreams" thing since I've had multiple really vivid dreams where I've read like entire pages of things from the internet. One of the dreams I even read this very wordy comic twice over and it was the same each time
AI trying to produce text is like our brains trying to process text in our dreams.
what is AI but the dreams of a million scientists
The only text I clearly remember from a dream is a candy bar labeled "Carraheers" almost certainly prompted by me (while awake) reading a candy bar label and wondering what carrageenan was. It was in the style of a Three Musketeers bar by the way, making the dream word a portmanteau of carrageenan and Musketeer.
Now that I think about it, that's eerily similar to what these programs are doing...
I don't know if i'm weird, but i remember any text in my dreams being completely coherrent and sensible
Can people stop making this comparison please
@@rowboat10 No because it's a good one.
My PFP is made by ai! I was asking for "Himbo" in white on a space background. I'm very happy to see text that's so close! (This was not my first try though haha)
that pfp looks so cool yet so cursed.
i guess it worked because before reading the whole comment I check your pfp as you mentioned it and I thought it looked like "himbo"
for parsing, stable diffusion REALLY loves proper commas, especially the oxford-comma, and a list of properly-sorted adjectives. Stable-diffusion will take ANY metaphor literal, and metaphors are mostly about adpositions (abstractions around adpositions, you can easily sort 10000 metaphors+idioms by the adpositions in them), so you should avoid adpositions/prepositions (they may just imply a metaphor that you may not even realize that it could be taken too literal) and use adjectives for context instead, and use directions+containers as adjectives, instead of adpositions, and sort adjectives properly!!!
For those not familiar with “adpositions” like me until a few seconds ago, according to Wikipedia, “Prepositions and postpositions, together called adpositions, are a class of words used to express spatial or temporal relations or mark various semantic roles. A preposition or postposition typically combines with a noun phrase, this being called its complement, or sometimes object.” From what I can tell, especially in the context of prompts for AI, they are a class of filler words I have generally learned to exclude despite not knowing it, just because as ollllj said for an AI they are often filler that can introduce lexical ambiguity.
5:44 Yard Yay just oozes positive energy, really puts me in shopping mood
5:13 was that a welsh ll?
Oh, cheese, how you bring me such joy
Your flavors, textures, oh boy!
From cheddar to brie, you never disappoint
A perfect addition to any savory joint
You melt so beautifully, oh so smooth
On pizza, in mac and cheese, you always improve
Your sharpness and tanginess, a delight to the tongue
I could eat you all day, until my belt is undone
From grilled cheese sandwiches to cheeseburgers galore
You elevate every dish, leaving me wanting more
Oh, cheese, how I love thee, let me count the ways
Forever in my heart, always a place to stay.
There are a bunch of little tricks you can do with your prompts in order to get better results, for example: Dall-e allows you to apply weight to certain elements of the prompt like so:
"[a cat::8] in [a field of corn::2]"
The cat has a lot more weight and as such will be much more prominent in the image, and there will be a hint of a corn field, probably in the background. If you were to swap the weights, the field becomes the more important element and the cat should be almost a background element.
I think you're talking about Midjourney, not DALL-E.
@@rkvkydqf you're probably right. It's been quite a while since I last played with these AI image generators.
7:00 it reminds me of those videos that show how English sounds to people who don’t speak it. It’s got all the proper sounds, but without any discernible meaning.
this is really interesting. I remember playing around withone of those ai image generation sites maybe 6 months ago. asking for "motivational poster" or "motivational quotes" and getting what looks like stylistic typography but the text itself was foreign and alien looking. very cool to see that they have leveled up to actually generate english characters, even if the words made with them are incomprehensible
I always find AI typography to look like what a person would create in a script they don’t read or write, whilst looking at examples. I’d imagine my attempts at drawing a page of Japanese characters or Arabic lettering would look just about as convincing and read as the same level of gibberish.
Also, no interest in trying MidJourney? It’s the most “artistic” text to image model at the moment, easily surpassing many junior-level illustrators (though it tends to create images of a very distinct style one learns to recognize easily after some practice)
Midjourney is still in beta. I had a chance to try it out and it is great for more artistic prompts but I don't think it would get great results for a video like this.
@@EvanBear It honestly can! You sometimes have to run a few prompts but it makes amazing things, in an artistic style. It's great with fantasy concepts, DnD characters, magic or pretty lighting.
@@OhSoUnicornly Yes it's very artistic but it doesn't do well with "in the style of X" type of prompts. It has its own style. I got amazing cyberpunk images out of it tho, so I'm not knocking down Midjourney as a tool but if you still have attempts try the squirrel on a red table eating colorful marbles prompt.
@@EvanBear Ah I see - it would do great with "strange animal in a field" type prompts though, which Mike seemed to be lamenting the lack of in this one!
Fun fact, the type of text you are referring to is called "greeking".
For the purpose of disproving the stock image theory, I'm surprised you didn't ask AI to draw an Atomic Shrimp.... The bigger question is who owns the IP / copyright rights to the output images? Is it the person who generates the image?
I guess it depends on uniqueness and the transformative nature of the result, but I'm not sure if that's quantifiable. Generally you are allowed to create a derivative of the original work if it doesn't compete with it.
The USCO (U.S. Copyright Office) just decided that images generated by an AI cannot receive copyright protection. Here it were pictures for a comic generated by Midjourney. To summarize very shortly, the text prompt doesn't influence the result sufficiently to be considered the source of the picture, and only things created by a human can receive copyright protection. This means pictures generated by an AI are in the public domain.
There has to be creativity and legal precedent at least in the US says that has to be human. PETA tried to claim copyright for a monkey which had taken its own photo by means a movement sensor. The court agreed the human photographer had not taken the photo so he lost his copyright but the monkey couldn’t be given it so the photo is now in the public domain. I suspect any AI generated material would fall into the same category
@@molybdomancer195what about if it's an AI generated artwork of a copyrighted character? I remember a month or so ago there was this whole big issue of vtubers telling their fans to only tag them in art they did in the traditional way and being against AI art. In that instance I kind of agree with them because the AI is sampling artwork created by other people and thus it's not entirely original.
In the US, no one owns the output and the images cannot be copyrighted (transformative media, fair use)
I assume it’s only a matter of time before corporations lobby to get ownership of AI images so their precious money-saving AI-made brands can be monetized to the moon and back
6:47 sweet, þ representation
I’m glad Simon was a great sport about it! I love both of your respective channels and enjoy their content very much!
Mike, I take slight issue with your interpretation of the AI's interpretation. You've gone to great lengths to clarify that the AI is not sentient and is not building things out of thin air, but on things it was trained with. Therein lies the issue, these massive training datasets were for the most part copyrighted images that were ripped whole cloth from image scrapers. Some of which already came tagged with what they were, others were tagged manually in vast tagging farms in less developed countries. I wish that you'd keep that in mind rather than this middle ground of "well the AI 'knows' things and puts them together." Otherwise, thanks for the great video as usual, it is very funny to see the AI break in those ways.
That's an interesting discussion for sure, but you're not disagreeing with anything I said in this video
@@AtomicShrimp No, it's more of, "I wish there was an asterisk here" rather than disagreement.
@@saberleo4567 Sure, the ethics of data collection for training these things is a big mess right now. My own view on it does not align with either of the extremes.
To elaborate: I think it's entirely reasonable for artists to find it objectionable to their work being included in the training data. I'm not completely sure if such objections should be absolutely entertained - because in a sense, I see an analogy between me studying a painting in a gallery and the algorithm being trained on an image, except no analogies are ever exact - image generation algorithms are optimised for greater competence in reproduction than I can achieve using my own skillset, so I cannot mimic what I have seen as well as an AI can. In addition to all of this, if the western world decides it's not OK for copyrighted works to be included in AI training data, other parts of the world with a more laissez-faire attitude to intellectual property rights won't care much, and will take advantage (this is an appeal to consequences, but it's also a real concern with precedents - look at the Chinese knock-off market). In summary, my own thoughts on this matter are a big unresolved mess. I can see virtue and fallacy in both 'sides' of the argument being argued.
@@AtomicShrimp I understand that side, sorry that I glibly painted your stance as unreasonably centrist, so far I'm reasonably satisfied with the conclusion of the US Copyright Office that AI art is not in it of itself copyrightable. I think that that is the best way for AI art to spread as a tool for assisting humanity rather than as a tool to subjugate it.
1:52 What's really eerie is the reflection of the rest of the room, unseen in the direct view.
8:41 "for my hair is bald" such an enlightening end to a poem about cheese 😭
Something to remember as well is that doesnt just have access to English text, so it's probably mixing a lot of languages into one when it tried to 'create text'
I'm sure an overwhelming majority of its database is english text though
4:42 these pronounciations kills me 💀💀💀
You and Simon Roper, this is gold! Thank you for this, I didn't know that this is what I needed.
I want a print of that extended Jabba wocky picture it is great😜
Me too
I wonder if the pills in the image at 2:05 are the codes that will change the color of the mouse.
5:33 in the second three text says "Lor", which in goofy ancients language from kirby means paradise.
A couple of moments went by where I was thinking "has he multi-layered this by using an AI to read the script in his voice"? That'd be amazing.
I have thought about looking into voice cloning, and I have seen some impressive demonstrations of it, but they are never nuanced enough for me to hand over a script to them. Sometimes I don't know how I am going to express a particular phrase until the words come out of my mouth
8:30 sounds like Dutch or some other broken half english half german mix like Irish
with a sprinkle of French
4:04 crossover episode??
I wonder what could happen if you asked the AI to make a picture of the alphabet. Or maybe ask it to create its own language? Thank you for the video!
PLS NO
ONE GUY DID IT AND I WAS TERRIFIED
My favourite UA-cam channel Mike, thank you! Love the variety and especially the shorts. I'm just hoping we might get more positive news on poor Davis Bon soon...!
I love it when two of my favourite channels unexpectedlöy clash.
Yes llol ythlet lor
Itt ettpy gtret
That typo wasn't even deliberate. My keyboard is acting up at the moment.
1:42 DALL.E DIDN'T NEED TO DRAW A YOUNG MOZART JFC-
I saw Pergency Wanged on the Pyramid Stage at Glasto in 2018.
I think the only point of AI that I've personally found, is the innate experience of finding its limits. Through your own journey to the edge of a model, you'll discover more about how your own mind works, and your position in the whole of human existence.
And it's damn scary sometimes :)
I'm also just now realizing this is the exact same meaning that humanity might have found at the top of a mountain. Or the bottom of the Mariana Trench.
I've personally climbed down and up the grand canyon in a single day. And that feeling has real comparisons to my night of 'texting' chat gpt
We're in a very strange place as a species.
Entire leaps of societal progress are just hopscotch to train the future.