My First Impressions of DALL·E from OpenAI

The powerful text-based generator can create images in almost any style imaginable

I’ve been playing around with AI image generators for a little while now. My first foray into this artificially generated world was Deep Dream Generator, which uses existing images to create something unique.

While I thought that was cool, my mind has officially been blown with DALL·E from OpenAI, which only requires a text description to generate a picture.

I had heard some rumblings about DALL·E, and had seen some impressive samples, but I was still skeptical. Artwork created from just a few words? Sure. Whatever. So I joined the waitlist for DALL·E on August 21, and was accepted into the platform on August 28. I thought the wait would be longer, but I’m definitely not complaining.

I jumped right into it and started creating. My first image was based on a photo I had taken the day before of two random balloons. I had imagined a female model holding the balloons in the middle of the street. So I typed this descriptor into DALL·E:

impressionist painting of a beautiful woman with long black hair holding two balloons one is blue and one is green in each hand in the middle of a city.

This is what the GPT-3 trained image generator came up with:

It looks like something that could be hanging in a gallery, in my opinion.

I then tried some other descriptors that randomly popped into my head, like this one:

digital art of a muscular man is being chased through the mountains by a huge frizzly bear with a camera.

Yes, apparently I called it a “frizzly” bear by accident, and my sentence barely makes sense. It’s not what I had envisioned, but what DALL·E generated was even better. The scale of the bear, the beautiful mountainous backdrop, the bird, and the interesting patterns work well together.

Young imaginations yield big results

My eight-year-old son then got into the action. He loves sea monsters, so he simply asked DALL·E for a giant octopus attacking pirate ship. His eyes got really wide when he saw the results:

My son, with a very vivid imagination fuelled by Minecraft, then asked for this:

pixelated skeleton standing in a dark room next to a glowing reactor covered is a strange pink moss

Again, we were both impressed with the results.

I excitedly took over again. I wanted to create a “fantasy” world type image with a dragon. This is where I confirmed that the order of wording is very important to DALL·E. I wanted the human figure to be holding the sword:

fantasy world with a dragon with fire coming out of its mouth fighting a man with a sword

Yes, the dragon is holding the sword, as if it didn’t already have a massive advantage.

Any artistic style you can dream of

You can ask DALL·E to generate images as digital art, an oil painting, and so on. For example, I tried creating a self-portrait in the style of a pencil drawing, and the program generated something freakishly akin to my own drawing style without a visual reference:

(I was a bit confused by what looks like a very long nose hair. But I believe that is where the AI pencil started from, and then it traced its way back out of the page.)

I then used a pop culture reference in my text descriptor, in this case an evil McDonald’s logo with scary teeth. DALL·E did not disappoint. (However, I’m not really sure what MIDECALIS is supposed to mean. Perhaps this is what they call McDonald’s in an alternate universe?)

I also tried generating something that my spouse might be able to use as a promo for her Pilates classes:

impressionist woman in purple pants doing pilates in a sunlit yoga studio

I’ve done a “bridge” in Pilates before, but never supporting my entire weight with only my head. I’m sure it can be done, but not at my fitness level.

My spouse believes the AI might be confused by the millions of unrealistic body images on the internet. In other words, this is what DALL·E thinks is a normal body type and pose. (Notice how I didn’t describe the woman’s features to the image generator — this was chosen as the default.)

AI images are here to stay — and will only get better

At this point, I am hooked on DALL·E like someone who just took a free sample from a dealer. DALL·E gave me 50 free credits when I joined. I’m already down to eight credits, but I can buy more or wait for 15 more free credits arriving on September 29.

It costs $15 USD to buy 115 credits, which I think is very reasonable considering you can use these images for commercial purposes according to the terms of use. You can even sell your “Generations”. (On a side note, I tried pushing the limits into more mature themes. It seems DALL·E is a bit of a prude as per its content policy.)

DALL·E generates multiple results from each description, all of them quite different. You can save the images directly to your computer as a PNG file, at 1024x resolution, which is more than good enough for the web. I also like how the image filenames default to your text descriptor, which is useful for SEO.

Obviously, there is huge potential here. DALL·E exceeded my expectations. It truly is a next-generation AI. I can see this becoming more advanced to the point where faces look flawless, which will expand the commercial uses. And since DALL·E can generate realistic images, I don’t see why a future version won’t be able to generate video clips based on simple text prompts.

It’s worth getting on the waiting list and playing around with DALL·E from OpenAI. I don’t think traditional artists should worry (yet), as the application still requires a human imagination to function. (The same goes for AI writing.) There’s also some fine-tuning involved if you were using the images for specific purposes (such as branding.)

However, there’s no denying that DALL·E from OpenAI is definitely one giant leap for machinekind.