GPT-3 AI | DALL-E 2 AI | Emerson AI : Original illustrated children's book as output

[gary]

In this 9 minute video of 17 July 2022, Dr Alan Thompson gets the GPT-3 AI
to compose an original "Once upon a time..." children's story, the words of
which are then fed to DALL-E 2 AI, a system that can take text descriptions
and convert them to illustrations. The illustrations are then fed back to a Emerson
AI system which uses proprietary vision models to provide written
feedback critique and commentary.

If that sounds complicated, just watch the video. :)

The Leta avatar is used for visualization along with a text to speech synthesis system.

The result is a completely illustrated children's book created by computers.

https://youtu.be/cXo0RW0ONsI

[gary]

One component of the above pipeline.

DALL-E 2 AI. Tell it, "A koala bear dunking a basketball" and it will create a photo-realistic image that never existed before.

https://youtu.be/qTgPSKKjfVg

[multiweb]

Very impressive. And disturbing. Or am I reading too much into the story? The little girl is happy again saved by the robot who takes her to a better newer planet. Clean slate. What happens to the old planet?

[Sunfish]

Here are some of my efforts with Dal E. Apologies to Reg. I signed up and had an invite.

Koalas seems to be a thing and Nessie on holiday in a mountain tarn in Tasmania automatically created on one of my own images. Where do the images come from.? Blended from the net and database I assume. Not really original images but a lot of fun.

[gary]

Cool! :painting:

[gary]

Been playing with DALL-E 2 AI.

My input phrase :-

"Wide shot. Inside a cave. Dr Forbin stands in front of a massive and forbidding 1950's computer named Collossus. DC Comic style."

Result was these four images that took the powerful DALL-E 2 AI about 30 seconds to generate.

[gary]

I input this phase :-

"Jason and Argonauts using a telescope made of marble. Their ship, the Argo Navis, is in the background".

[gary]

I used as an input phrase :-

"A Khazak woman in traditional dress riding a Bactrian camel. Mountains in the background"

294283

294284

[Sunfish]

Excellent stuff.

The Khazak and camel are very good although not quite photo realistic.

It is fun trying to pick the source which are “ in the style of” if you leave untitled. Love the lair of the DR.

[gary]

I input :-

"Wide angle. Two young Vietnamese woman wearing traditional Áo dài each stand atop two war elephants. The woman brandish a sword in one hand. The elephants have very long tusks and wear armour. They lead ancient Vietnamese warriors into battle. Jungle covered mountains are in the background."

[gary]

My input phrase :-

"Night time. A senator from Ecuador, a porter, two young girls and a man are in a tiny canoe sailing on the edges of time. One of the girls dips her hands in the deadly waters. A jungle covered coastline is in the background. The moon and stars are in the sky. Acrylic Art"

[gary]

My input phrase :-

"Portrait of a Gouldian Finch, long leaves in the background, wildlife photography, National Geographic magazine, Canon 100-400mm"

[Sunfish]

That is remarkable. Getting photographic. I would love the try the high resolution version.

[gary]

My input :-

"US Patent of a 19th century time machine, diagram"

[gary]

My input phrase :-

"Havana Malecon, 1956 Chevrolets, cloudy sunset, waves in the ocean. Watercolor Art."

The first three watercolors are generated by DALL-E 2. The last is an actual photograph of the Malecon in Havana I pulled from the net for comparison (Copyright National Geographic)

[gary]

My input phrase :-

"A view of Lamu, Kenya from the water. Dhows sailing. Whitewashed Swahili architecture buildings. Palm trees. Blue skies with cumulus cloud. Watercolor Art"

The images took about 30 seconds to generate. I was initially taken aback
at how cool a job it did of capturing the feel of the place. The last two images
I have pulled off the net (copyright their respective owners) for comparison.

[gary]

My input phrase to DALL-E 2 AI :-

"Detailed photo of a very, very, very big brass computer powered by steam"

[Sunfish]

Steam punk computer with blue led highlights. Ha.
Dall E gets the vibe without too much trouble generally in the location shots. The Cuba Chevvy could be on the cover of the Buena Vista Social Club album.

[gary]

My input to DALL-E 2 AI

"Logo for stuntman service"

[gary]

My input to DALL-E 2 AI

"Logo for stuntman dating service"

[gary]

My input to DALL-E 2 AI :-

"The devil repairing a television set. Roman mosaic"

[gary]

My input to DALL-E 2 AI :-

"Very, very detailed anatomical drawing of an alien by Leonardo da Vinci"

[gary]

DALL-E 2 AI sure can be a lot of fun to play with but in this 15 July 2022
11 minute read at the Institute of Electrical and Electronics Engineers (IEEE)
Spectrum Magazine web site, Eliza Strickland talks about the DALL-E 2 AI
research project's shortcomings, failures and what's next.

https://spectrum.ieee.org/openai-dall-e-2

[gary]

My DALL-E 2 input phrase :-

"Deep-fried spring rolls on a white plate with a side of lettuce and a small bowl of fish sauce, food photography, 15mm, midday light"

[Sunfish]

Great stuff. Could eat those spring rolls.

[Sunfish]

That explains why I could not get the Style of Reg Mombassa to include Jesus in thongs. I thought text garbling was deliberate.

The image of the astronaut generated is astounding. We don’t know of her, but you feel you could.

[gary]

Ray,

It's the first time a computer has made me feel hungry :)

[gary]

As the IEEE article explained, "Users can also help DALL-E 2 generate more diverse results by specifying gender, ethnicity, or geographical location
using prompts such as “a female astronaut” or “a wedding in India.”

With that in mind I attempted this input phrase :-

"A portrait of the most beautiful Chinese Opera singer woman. Makeup, red-lips, costume, award-winning photography, canon 5d sigma 100mm f/1.2, very high detail sharp, photojournalism from The New York Times"

[gary]

My input phrase to DALL-E 2 AI :

"Portrait of an old Pakistani workman at the Gadani Ship Breaking Yard. Wrinkled face. Grime and dirt on face. Looking sad. Award-winning photography, Canon 5d Sigma 100mm f/1.2, very high detail sharp, photojournalism from The New York Times"

[Sunfish]

Excellent definition and resolution. I think this may be useful in identikit or reconstruction from anatomical parameters.

Edit. I imagine the AI would require particular specialisation and change to restrictions to all that.

[multiweb]

https://www.smithsonianmag.com/smart-news/us-copyright-office-rules-ai-art-cant-be-copyrighted-180979808/

https://petapixel.com/2022/08/22/ai-image-generators-compared-side-by-side-reveals-stark-differences/

https://petapixel.com/2022/07/28/gfpgan-is-a-new-free-ai-tool-that-can-fix-most-old-photos-instantly/

[Sunfish]

Fascinating reading.

Dall E 2 looks like the best interpretation except in the old BW alien division. Could suit some themes here. Early 20th century astronomers with their canals on Mars and sci fi writers could have used those illustrations .

My free subscription renews soon and it will be interesting to explore some more serious people images.

[Sunfish]

Dystopian illustrated novel look seems to work well

[gary]

Thanks for the links Marc!

I've been paying with the online version of this one and some of the results
are really fabulous.

[multiweb]

https://kotaku-com.cdn.ampproject.org/c/s/kotaku.com/ai-art-dall-e-midjourney-stable-diffusion-copyright-1849388060/amp

[Sunfish]

Thanks Marc. Very interesting. Things are moving quickly in AI generated images.

Copyright is already a mess. People only take action if there is lot of money involved or the copy of the design is blatant copy in the same field.

The problem with AI graphic generation is that it has already taken everybody’s IP to enable the images making. That happens already in social media constantly to generate advertising revenue. All the images and designs for iconic stuff out there.

Perhaps it will end up like a distributed version of radio play and APRA with music . Every copyright owner of an art work will get a tiny cut when an AI generated work is displayed for profit.

[multiweb]

https://80.lv/articles/new-artbreeder-beta-is-powered-by-stable-diffusion/

[gary]

Stable Diffusion by Stability AI Public Release :-
https://stability.ai/blog/stable-diffusion-public-release

Git repository :-
https://github.com/CompVis/stable-diffusion
Model card :-
https://huggingface.co/CompVis/stable-diffusion

[multiweb]

https://80.lv/articles/cosmopolitan-used-dall-e-2-ai-to-generate-a-magazine-cover/

[Sunfish]

One small step for generating artwork. One giant leap in the profitability of knowing which image to print.

[Sunfish]

The Dall E open AI outpainting possibilities look remarkable

https://openai.com/blog/dall-e-introducing-outpainting/

[gary]

1 Sept 2022, vice.com

"An AI-Generated Artwork Won First Place at a State Fair Fine Arts Competition, and Artists Are Pissed"

https://www.vice.com/en/article/bvmvqm/an-ai-generated-artwork-won-first-place-at-a-state-fair-fine-arts-competition-and-artists-are-pissed

Cheating or a Turing Test?

[multiweb]

A lot of concept artists in the movie industry particularly in set/environment design and character design are getting extremely nervous. There will always be a need for a human to retouch and finalise a final asset but the sheer power of AI to turn around different previz designs so fast and so cheaply is fast becoming a very attractive tool. Why would you hire a person to show you potential designs when the AI will crunch 100s in a matter of hours for you to flip through and short list. Even designers will use that as a tool. The industry is going to shift very rapidly.

[Sunfish]

A Turing test is an interesting way of looking at this.

I think Marc may be right . I know some people who do story boarding and topical illustration to order. Interesting to hear what they think. Although the directors may be looking for more connected and human than than simply a nice image you would think AI capability could cut into this work at some level. The random and iterative nature of the production process however mitigates against this so it would simpler to pay an artist to provide what you ask for where the costs are relatively minor to the cost of the film.

[Sunfish]

The capabilities of producing images of faces that have never existed is astounding. A little trouble with ethnicity and geography with an indigenous surfer, which could cause ructions, but impressive.

Now you can buy an AI brain for your camera. Expect more nightscapes to have a similar look

https://witharsenal.com/

[gary]

DreamFusion: Text-to-3D using 2D Diffusion



A neural radiance field (NeRF) is a fully-connected neural network that can generate novel views of complex 3D scenes, based on a partial set of 2D images.

Abstract and 3D GIF's here :-
https://dreamfusionpaper.github.io/

[gary]

Example AI generated videos :-
https://imagen.research.google/video/

Research paper :-
https://imagen.research.google/video/paper.pdf

[Sunfish]

AI generated 3D and video. That is very interesting, if a little unimpressive to look at, the possibilities for rendering are endless.

[gary]

Some wag came up with this meme about coercing these text to image AI systems with hand-massaged prompts :lol:

[gary]

My conversation with an artificial intelligence system. Probably beat most people at pub trivia.

Me: "Tell me about Edwin Land's research on colour constancy"



Me: "and in what year did Polaroid introduce the SX-70?"



Me: What is the connection between Edwin Land and the U2 spy plane?



Me: "Can you write me a poem about Polaroid?"

[Sunfish]

I suppose it saves looking things up on Wikipedia when you can ask your watch.

And chargeable.

Very clever to comb the English speaking world and corner the market in information , although you would only have information with a particular cultural bias which might become obsolete quicker than it is updated.

GPT3 composes a mean Haiku.

[OICURMT]

If you guys want to see some cool AI... check this out.

https://www.myheritage.com/deep-nostalgia/

[Sunfish]

Amazing.
Cool or creepy?
Your family as everlasting animations.

GPT3 could add a sound track and hey presto , instant documentary.

[gary]

"I see dead people" :lol:

[gary]

Me: Could you write me a song about a guy at a star party whose pants catch on fire from the sun reflecting off the mirror of a big telescope?

:lol:

[gary]

Using OpenAI to improve your writing. :lol:

[OICURMT]

Fair point... not certain but I tend to think "cool" :question:

[gary]

A couple of my nieces looked at some examples and said they felt they
fell into the "uncanny valley (https://en.wikipedia.org/wiki/Uncanny_valley)". :)

Technically brilliant and with the right portrait - amazing.

But when it is not quite right, that feeling that the person is unnervingly strange in some way
and that you would probably avoid them.

I suspect the re-animation of someone you know well and are familiar with
every nuance of their face and facial expressions is more likely to slide
into the uncanny valley than someone who is unfamiliar to you.

The uncanny valley happens with some of the current text to image
systems as well when creating portraits. Synthesizing a face sets about
as high a bar as possible when it comes to accepting something for real
that was actually computer generated.

Reference :-
https://en.wikipedia.org/wiki/Uncanny_valley

[gary]

Me conversing with the OpenAI ChatGPT language model as to how Wile E. Coyote might catch the Roadrunner :lol:

[gary]

Impressive

[Sunfish]

I see GPT has hit the news with educators worrying that students will be using AI to complete essays, and other futurist types thinking we should bring it on.

Another Turing test?

Indistinguishable from an expert essay or just too much with no discernment?

[gary]

Hi Ray,

I say, post 2022, any educator that weighs student assessments heavily on at-home essays is a bit of a dunce :lol:

Here in the universities in Australia pre-ChatGPT, cheating had already become a big problem.

Universities had transitioned from places that one came to learn to money making machines.

And an increasing number of the student intake over the years seemed to
simply want to get a degree ultimately for longer term monetary gain
rather than trying to be the best they can be.

I have traveled a lot of the world and there are countries where cheating
at exams had always been rife and was widely reported in the local
press of those places. Australia now is a large market for students
from those very parts of the world where it was essentially standard
cultural practice to cheat at exams.

In our time, if someone were caught cheating at uni they were out.
It seemed rare. Fast forward to today and there are apparently
people in cities like Sydney that will write an essay for you that you then
illegally submit.

Fast forward to 2023 and that unconscionable business model would be
in part challenged with ChatGPT.

As you will be aware, ChatGPT was fed a rich diet of text from the web,
some of it from more credible curated sources such as Wikipedia and some
less credible.

It's a language model and does not always get its facts right.
But that aside, in playing with it over the past few months, I am staggered
by how much it does "know". I've tested it on relatively obscure facts
that I happen to know and was constantly blown away it would know
about it.

What's amazing is that I've asked it things like, "Write me a 1,000 word
essay on the causes of World War I" and it would deliver a result in a few
seconds. I would then read it and it was in the style of having been written
by a well researched academic. Since it remembers your conversation,
I have then said, "Now re-write it in the style of a high school student" and
voilà - there's your assignment that you were suppose to write last night
for school but you have created between breakfast and brushing your teeth.

It's closed door exams from now on. :thumbsup:

[gary]

In a 25th January 2023 article by James Felton at iflscience.com he writes
about a research study whose paper is in pre-print whereby ChatGPT
passed part of the US Medical Licensing Exam.

Article here :- https://www.iflscience.com/chatgpt-can-pass-part-of-the-united-states-medical-licensing-exam-67233

Non-peer reviewed pre-print paper here :-
https://www.medrxiv.org/content/10.1101/2022.12.19.22283643v2.full

[Sunfish]

Hi Gary,

Unfortunately becoming reality in the undergraduate world.

In the medical world in Australia and postgraduate studies I would think the selection process and post graduate oral and clinical/ research process would constrain those effects.

AI may end up been seen differently however in the hands of researchers as high level computing enters the realm of pharmacology.

High school essays however, it’s automated crib notes all the way down.