Michelle Brändle
Background information

From horsies to horror – creating fantasy landscapes with AI system DALL-E

In addition to ChatGPT, OpenAI also does images. A software called DALL-E is can now modify images by means of markers and descriptions.

With DALL-E, you can create images via the paid version of ChatGPT. Now, the images made by the AI system can be edited further. Not everything runs smoothly yet, so you’ll need time and patience to achieve the results you want.

In ChatGPT, the image generator called DALL-E is on the left.
In ChatGPT, the image generator called DALL-E is on the left.
Source: Michelle Brändle

Operating the interface is easy. You can either tell DALL-E what you want it to change by writing a description or prompt in the chat window. Or you can use the newly added editing tool by clicking on the paintbrush icon at the top of the chat window. The latter gives you three options:

  • adding something to the selected area/object
  • removing selected objects
  • editing the selected area

Editing only works with images created by DALL-E. However, you can also upload an image for DALL-E to make its own version. I create a template myself using the image AI.

Using the following prompt, DALL-E generates an illustration for me within one minute:

Generate a fairy world in pink with unicorns, butterflies and a castle and lots of flowers

DALL-E responds with a longer text below the image I’ve created:

«Here’s the fairy world you envisioned, brought to life in hues of pink with majestic unicorns, delicate butterflies, a whimsical castle, and a sea of flowers. I hope this matches the magical scene you had in mind! If you’d like any changes or additional elements, feel free to let me know.»

The image is overly pink for my taste and a bit heavy on the butterflies. They need to go. To erase them, I use the selection tool, mark some butterflies and tell the AI to remove them.

My modification involves marking the butterflies I want to eradicate.
My modification involves marking the butterflies I want to eradicate.
Source: Michelle Brändle

DALL-E needs two attempts. The first time, it reduces the size of the butterflies instead of getting rid of them. It takes about half a minute to calculate each change.

It spits out the result I wanted after the second attempt. The butterflies have been removed.
It spits out the result I wanted after the second attempt. The butterflies have been removed.
Source: Michelle Brändle

Next, I tell DALL-E in one sentence what changes to the colours I’d like. More colours, less pink. Interestingly, the AI doesn’t just change the colours, but the entire image:

After the colour changes I wanted, the image is no longer the same. Now I don’t like it any more.
After the colour changes I wanted, the image is no longer the same. Now I don’t like it any more.
Source: Michelle Brändle

The castle and unicorns are different. That’s a pity and wasn’t the plan. Even the image format isn’t as wide as it was. I’ve had enough of all this gaudiness and make a radical adjustment to dark horror:

DALL-E completely reworks my original idea and makes a darker version of my fairyland.
DALL-E completely reworks my original idea and makes a darker version of my fairyland.
Source: Michelle Brändle

Again, DALL-E makes changes to the whole image. Nothing’s as it was. Several attempts bring unexpected changes, such as a dark army of fairies in front of the castle. At some point I’m happy with the result: grim skeleton unicorns. Perfect.

What if I combine the dark and the cheesy ideas? I tell the AI to pop a pink unicorn in the middle and mark the exact spot I want it:

Getting the command right for a realistic, pink unicorn to be placed in the empty space is harder than expected.
Getting the command right for a realistic, pink unicorn to be placed in the empty space is harder than expected.
Source: Michelle Brändle

The result’s a soulless unicorn-shaped blob of colour. Even after several modifying attempts, it’s just not working. The space stays empty when I ask DALL-E to generate a realistic-looking unicorn. Strictly speaking, that’s not even wrong. After all, a realistic unicorn equals no unicorn. Rewording the prompt doesn’t do the trick either. The space stays empty. I tell DALL-E to insert more unicorns. Again, it doesn’t do what I want it to. It’s probably best to stop before I or DALL-E start making things even worse.

Several attempts in and I’ve only made the horror landscape worse.
Several attempts in and I’ve only made the horror landscape worse.
Source: Michelle Brändle

Verdict: inspiring, but not the full package

At first glance, customising images to your taste with the help of artificial intelligence DALL-E seems handy. However, the system has several flaws. If you want to remove an object from an image, the AI often replaces it with a new object. If you’re making changes with the editing tools, the AI modifies the entire image. For example, a castle will look different after a colour adjustment. But with a little patience, you can still get alright pictures. Albeit with unexpected components.

DALL-E’s new editing function hasn’t changed my view on image AI. I still find them immature tools that serve as a source of inspiration for your own images. And in my view, an image AI shouldn’t cater for more than that.


Check out other AI tools I’ve tested:

  • Background information

    My Digitec summer hit 2024 thanks to the Suno AI programme

    by Michelle Brändle

  • News + Trends

    Do it with a doodle: AI tool "Stable Doodle"

    by Michelle Brändle

  • News + Trends

    Chatbot with a difference: dork GPT as a cheeky conversation partner

    by Michelle Brändle

  • Background information

    Photoshop: Adobe’s new AI image generator can do a lot – but it can’t do everything well

    by Michelle Brändle

Header image: Michelle Brändle

33 people like this article


These articles might also interest you

  • Background information

    Photoshop: Adobe’s new AI image generator can do a lot – but it can’t do everything well

    by Michelle Brändle

  • Background information

    StableDiffusion: Images from text input

    by David Lee

  • Background information

    "Adobe Firefly": The web app now generates AI videos

    by Michelle Brändle

Comments

Avatar