DALL·E 3 is here.

And it will change the way we prompt.

Created with DALL-E 3

When 2021 OpenAI presented DALL-E, it was the beginning of a new era. Digital art had been advancing for decades, and AI art was already here. But creating images with words became a new paradigm for the entire Generative AI movement.

I wrote a lot about prompt design with DALL-E 2, but the main principle was that the prompt strategy was your skill — you had to know how to prompt to get a photorealistic image or illustration.

With DALL-E 3 everything changes again. What was accessible before (powerful diffusion models) becomes even more accessible:

since you speak to machine instead of prompting it.

At the moment (and this could change) it runs on chatGPT (coming in the next few months for Pro users) and (rumour has it) on Bing.

In short: to prompt chatGPT you must ask it to describe a picture — 
and chatGTP will write the proper prompts.

So how does this work in practice?

If we want DALL-E 3 to generate an image of an astronaut riding a horse (the iconic DALL-E 3 prompt along with the avocado chair), we have to ask chatGPT to write a prompt.

Step 1. We just write:

Step 2. Next, we will see how chatGPT-3 writes appropriate prompts.

Step 3. DALL-E 3 creates four images and writes a short description for each one (not to be confused with the prompt, which is more technical):

I like this illustrative one:

Step 4. We can go further and ask the system to create these images in a particular style:

Here we are:

Prompt: “Baroque oil painting presenting an astronaut in a richly decorated robe and periwig, mounted on a splendid horse with flowing mane. They are surrounded by opulent curtains and marble columns, with cherubs floating above. The backdrop transitions from an elaborate ballroom to a vast expanse of the universe.”

Can we reproduce an image (for example, using a seed)?

Anyway, just ask chatGPT.

Here are 4 Baroque pictures:

Now I would like to modify an image and ask chatGPT if this is possible.

You can use any of the 4 images to work with it — as you can see, it modified the image by adding a cat at my request.

But do images have seeds? We often use seeds in Midjourney (or even in the Latent Space related StyleGAN) to address exactly one specific image..

Let’s see, if chatGPT is just hallucinating or not. Reality check?

THIS!

You may find some small differences, but it is still the same image, the same composition, and the same style.

NOTE: The seed does not work across all DALL-E 3 generations. It only works in the same session. In other sessions, you will get different images with the same seed.

If you want to modify the same image with the same seed, and replace a cat with a dog (as you do in usual outpainting solutions), it won’t work exactly as you are used to:

It’s already another generated image. Same style, similar composition — but there is nothing wrong with DALL-E. It’s up to us, because we haven’t defined exactly where we want to make the change.

Prompt design is over?

Those of us who have been using prompt strategies since the very beginning might cry out: will prompt designer jobs be killed now that chatGPT is writing prompts for us?

NO.

Firstly, because we have new exciting ways of creating images — in a conversation with a machine.

Secondly, if you politely ask DALL-E 3 to use your exact prompt, it will!

So, Prompt Designers are saved — and all others are enabled to create more.

DALL-E 3 is a new communicative way of human-machine creative collaboration.

In my next explorations, I will challenge the creativity of this dynamic duett of chatGPT + DALL-E 3.

Stay Tuned.