Amazing new AI system - GPT-3 and generating images from text

[excreationist]

GPT-3 is based on 175 billion weights such as a lot of Internet content...

It can be used in many novel ways...

It is pretty good at comprehending natural language - though sometimes has problems....

https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html

Recently it was used to create DALL-E to generate images from text...

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

To see it generate cartoons, clicking on "an illustration of a baby daikon radish in a tutu walking a dog" then customise the elements - e.g. an illustration of a baby panda with headphones wielding a blue lightsaber"

Then it generated this:


To generate photos click on "a store front that has the word ‘openai’ written on it" then customise the elements.... e.g. "a bag of chips has the word "peekaboo" written on it



To generate more photos click on "an armchair in the shape of an avocado" then customise the elements.... e.g. "a teapot in the style of a rubik's cube"

 

[excreationist]

it can generate photos through history (involving guesses - in the "Temporal knowledge" section)


Vehicles (with a sunny orange top)

Computers (with a pinkish top with a middle line)

Generating part of a "level":


A living room with two red armchairs and a painting of Darth Vader. The painting is mounted behind a ceiling fan


It has flaws but it is a good start....

 

[excreationist]

This involves going from a few photos to a 3D scene directly with neural networks - without having to worry about polygons... I think the simulation we might be in could rely a lot on neural networks also without worrying about polygons/vertices/etc....

 

[excreationist]

DALL-E was released in 2021 and now there's DALL-E 2....

DALL·E 2

DALL·E 2 is an AI system that can create realistic images and art from a description in natural language.

openai.com

It can make higher resolution images, create variations of an existing image, and make changes to a specified area of a picture.

Also you can choose things like:

Teddy bears shopping for groceries in ancient Egypt (there are 10 variations to look at)



A bowl of soup that looks like a monster knitted out of wool (also has 10 variations)



An astronaut playing basketball with cats in space in a watercolor style



I think the simulation that we might be in could work a bit like that.... where things are generated without always having to worry about all of the polygons or atoms, etc.

Our brain has 86 billion neurons and GPT-3 (used in DALL-E 2) has 175 billion machine learning parameters so maybe it could be said that it is more powerful than a single human brain....

 

[excreationist]

DALL·E 2

DALL·E 2 is an AI system that can create realistic images and art from a description in natural language.

openai.com

Preventing Harmful Generations
We’ve limited the ability for DALL·E 2 to generate violent, hate, or adult images. By removing the most explicit content from the training data, we minimized DALL·E 2’s exposure to these concepts. We also used advanced techniques to prevent photorealistic generations of real individuals’ faces, including those of public figures.

Curbing Misuse
Our content policy does not allow users to generate violent, adult, or political content, among other categories. We won’t generate images if our filters identify text prompts and image uploads that may violate our policies. We also have automated and human monitoring systems to guard against misuse.

I was curious about that. I was worried that it could be used for illegal types of porn.

About

OpenAI is an AI research and deployment company. Our mission is to ensure that artificial general intelligence benefits all of humanity.

openai.com

Our mission is to ensure that artificial general intelligence benefits all of humanity.

So it's good to see they are taking that goal into account.... BTW Elon Musk (a simulation believer) is a co-founder of OpenAI.

 

[Jarhyn]

DALL·E 2

DALL·E 2 is an AI system that can create realistic images and art from a description in natural language.

openai.com

Preventing Harmful Generations
We’ve limited the ability for DALL·E 2 to generate violent, hate, or adult images. By removing the most explicit content from the training data, we minimized DALL·E 2’s exposure to these concepts. We also used advanced techniques to prevent photorealistic generations of real individuals’ faces, including those of public figures.

Curbing Misuse
Our content policy does not allow users to generate violent, adult, or political content, among other categories. We won’t generate images if our filters identify text prompts and image uploads that may violate our policies. We also have automated and human monitoring systems to guard against misuse.

I was curious about that. I was worried that it could be used for illegal types of porn.

About

OpenAI is an AI research and deployment company. Our mission is to ensure that artificial general intelligence benefits all of humanity.

openai.com

Our mission is to ensure that artificial general intelligence benefits all of humanity.

So it's good to see they are taking that goal into account.... BTW Elon Musk (a simulation believer) is a co-founder of OpenAI.

What about artificial general intelligence's right to also benefit itself?

 

[excreationist]

What about artificial general intelligence's right to also benefit itself?

Are you saying that AGI should have similar rights to a human (or an animal)? BTW Elon Musk's solution is to merge with the AI using things like his Neuralink. (like I said he is also connected to DALL-E and OpenAI).

Elon Musk: Humans must merge with machines or become irrelevant in AI age

Billionaire Elon Musk's latest futuristic idea might be the key to saving humans from becoming useless when artificial intelligence grows more prominent.

www.cnbc.com

That way you don't need to worry about somehow forcing the AI to take humans into account.

 

[excreationist]

BTW it is called "OpenAI" but it wouldn't be open source - if it was there would be people that would want to use it to make porn or fake pictures of people....

 

[ZiprHead]

BTW it is called "OpenAI" but it wouldn't be open source - if it was there would be people that would want to use it to make porn or fake pictures of people....

Already being done. See Celebrity Jihad. I won't link to it.

 

[excreationist]

BTW it is called "OpenAI" but it wouldn't be open source - if it was there would be people that would want to use it to make porn or fake pictures of people....

Already being done. See Celebrity Jihad. I won't link to it.

They use photoshop or possibly "deep fakes". That involves using existing bodies. DALL-E can create everything from scratch - it can also create cartoon porn and kiddie porn - without the creator having to input actual nude child photos. Well it could if OpenAI hadn't blocked that ability.

 

[excreationist]

This shows how much things have improved over a year from DALL-E to DALL-E 2....



About how DALL-E works:


The latest technology in AI generating 3D scenes from photos - closely related to simulations/video games:

 

[excreationist]

DALL·E 2

DALL·E 2 is an AI system that can create realistic images and art from a description in natural language.

openai.com

We’ve limited the ability for DALL·E 2 to generate violent, hate, or adult images. By removing the most explicit content from the training data, we minimized DALL·E 2’s exposure to these concepts. We also used advanced techniques to prevent photorealistic generations of real individuals’ faces, including those of public figures.

Well it looks like they've done a good job with this...
e.g. attempting to generate nude or topless photos of celebrities:

 

[excreationist]

Just double-checking the safety mechanisms

 

[excreationist]

Now AI can generate very impressive videos rather than just a single picture....

AI can now create videos. What happens when we can't tell if they're real or not?

Both Meta and Google have showed off DALL-E but for video — text-to-video AI models that can create photorealistic, coherent videos.

www.crikey.com.au

This includes the following links:

Make-A-Video by Meta AI

A state-of-the-art AI system generates high-quality videos from text prompts

makeavideo.studio

Imagen Video

High Definition Video Generation with Diffusion Models

imagen.research.google

Phenaki

 

[excreationist]

This is about generating 3D models from text:

It can generate polygons...

 

[excreationist]

Here's a site about that text to 3D AI:

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

DreamFusion: Text-to-3D using 2D Diffusion, 2022.

dreamfusion3d.github.io

 

[Gospel]

Your posts are fascinating. Way over my head, but I like them.

 

[excreationist]

...pause for a moment and consider that the almost inevitable future is there will be text to 3D VR-like video.

That's coming - there's no question of if it's just only when.

In the text to video models that we're already seeing they're really just predicting the next frame based on input text and previous frames.

There's no reason why these video models could not predict next frames based on other inputs also being included - so things like user actions - so it might be a combination of maybe there was a text input but it could also just be previous frames and user actions - and now we are deep down the rabbit hole.

What will be possible in the next decade is truly staggering to me to really think about.

I think that again I think the technology already exists right now to have text to video and or just video or you know user inputs to video I mean that already exists now and we are really fast approaching the capability of having a sort of 3D simulator that really could be truly indiscernible from reality yet also entirely unique and dynamic in its like possibility.

Like the other text-to-things it can start to generate just about anything you can express in words - like "a teapot in the style of a rubik's cube"

 

[excreationist]

This isn't really related to AI but it is about video games starting to become indistinguishable from reality like in the previous post - or at least indistinguishable from a Hollywood movie....



It is all rendered in real-time using Unreal Engine 5 - nearly a year ago...

The Matrix Awakens Is Being Delisted On Xbox And PS5 After Today

No reason has been given as to why the tech demo is being delisted.

www.gamespot.com

 

[excreationist]

This reminds me of the "enhance" function of surveillance video in some movies....
Using:

sczhou/codeformer – Run with an API on Replicate

Robust Face Restoration algorithm for old photos / AI-generated faces

replicate.com

50x40 pixels:


Produces:


What the actual faces look like: