OpenAI has its biggest week in 6 months

🏝 TheTechOasis 🏝

Breaking down the most advanced AI systems in the world to prepare you for your future.

5-minute weekly reads.

TLDR:

AI Research of the Week: OpenAI Releases DALL-E 3 and Changes Image Generation Forever

Leaders: Understanding ChatGPT’s Biggest Upgrade Ever, Building the Sensorial Computer

🤯 AI Research of the week 🤯

It’s pretty clear by now that OpenAI makes announcements in a big way, as this week has come packed with revelations from the star-studded AI company.

Among those, the most interesting by far is the release of DALL-E 3, the new version of their image-generating model.

However, the reason it’s interesting is not really the fact it generates better images… but how it does it.

This time, DALL-E 3 is built on top of ChatGPT, and that changes everything.

In fact, they have explicitly claimed that DALL-E 3 intends to be the nail in the coffin of what could become the shortest-living hot job in history, prompt engineering, begging the question:

Will AI eradicate AI jobs just as quickly as normal ones?

Source: OpenAI using DALL-E 3

Refining what state-of-the-art means

When working with current image generation models like MidJourney or Stable Diffusion, one has to deal with many problems.

Are you listening to me?

First and foremost, current image generators, models that take a text description and generate the image that portrays that description, are terrible at following instructions.

In most cases, they won’t respect structure, meaning that if you require a certain object in a certain part of the image, simply forget about it.

Yes, models like ControlNet allow us to condition a model to generate an image that resembles a certain structure or, for instance, a given scribble:

Unlike other image generators that start the denoising process from a purely random point, ControlNet helps you retain a certain structure by parting from a much less random and clearer image (top-left), so the image generation model has a much better idea what it needs to generate

But ControlNet is only great if you’re an artist, but those like me who aren’t, have to deal with the tendency of these models to obviate critical parts of your description and, most often, will require extensive tailoring of the prompt, a concept known as prompt engineering, to end up with the ideal image.

But lack of instruction following isn’t the greatest problem.

The Great Impersonator

Just like Ferdinand Waldo Demara became a legendary impostor portrayed by Tony Curtis in The Great Impostor, image generation models have had their fair share of notoriety in the press for imitating several artists who, put mildly, were not happy about it.

But imitated artists aren’t the only ones unhappy about these models, because unless you’re living under a rock, you have seen AI-generated photos of Donald Trump’s arrest such as the one below:

These images are funny but can scar the people portrayed heavily.

And considering how voice synthesizers are becoming scarily good at imitating the desired voice, one can easily…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.