AI Image Generation Algorithms - Breaking The Rules, Gently

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...

[davidfirth]

I think we'll look back fondly on this early era of ai art generation, with all its weird faults and quirks.

[HattmannenNilsson]

Bingus the boosing wockis is one of my favourite fairytale characters.

[Quinn-krcp]

Given how many misspellings of “yard sale” already exist, I wouldn’t be surprised if “YARLD SALE!” was a real one too

[neilomac]

I genuinely laughed like an idiot at 'Bingus the Boozing Wockis'. That definitely needs to be a thing.

[trickvro]

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.

[adam]

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".

[mystakn]

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.

[thetank]

5:56 I don’t know why but this really made me laugh. The deadpan “WANGED” just got me

[Jb]

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.

[davidmonroe]

Can we just take a moment to appreciate flawless pronunciation of Welsh double l with no pause or hesitation?

[thermonuclearwarhead]

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.

[PawsOnTheBalcony]

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 👍

[markozagar]

Bingus the Boozing Wockis sounds like something I would like to read, actually.

[rideronthedrumbeat]

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.

[revimfadli]

5:44 Yard Yay just oozes positive energy, really puts me in shopping mood

[SophiesLoaf]

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

[plasmasupremacy]

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.

[LukeAps]

This thing creates exactly the text of messages in my dreams. Hilarious!

[wariolandgoldpiramid]

AI trying to produce text is like our brains trying to process text in our dreams.

[JohnDlugosz]

1:52 What's really eerie is the reflection of the rest of the room, unseen in the direct view.