DALLE 2 – the OpenAI project that generates digital art from typed descriptions – is now open to all. Everyone can roll up and get 50 free AI-generated pieces of art.
To save you burning through your free credits by asking it to perform tasks it’s incapable of, or doesn’t handle very well, I’ve been testing the service to find out what it’s good at and where it strays off course. Here’s what I’ve found.
DALLE 2: What it’s good at
DALLE 2 really comes into its own when it’s asked to produce something abstract or artwork in a particular style. Think more magazine illustration than photo-realistic images.
You can see this from the examples that DALLE 2 offers up on its homepage. Tasks such as “An armchair in the shape of an avocado” or “A blue orange sliced in half laying on a blue floor in front of a blue wall” are perfectly rendered here.
Specificity is also a good thing when it comes to feeding DALLE-2. The more information you can give it, the better it performs. For example, I asked for “A football goal being scored by a chimpanzee in the style of a pop art painting” and got four brilliant options, including this:
Similarly, I asked for “a pencil drawing of a greyhound running into the sea” and this was one of four equally excellent results:
Telling DALLE 2 exactly what you want and the style of the image seems to deliver the best results.
DALLE 2: Where it goes wrong
Where Dall-E 2 seems to struggle most is when you’re asking it to deliver something photo-realistic or containing very specific, identifiable items. For example, I asked it to produce “a vicar reading PC Pro magazine” and this was the best of the four results:
The image isn’t terrible, but every time it tried to create the magazine it became a jumbled mess. You’ll note the vicar’s glasses are not quite right, either.
DALLE 2 is also clearly trying to avoid being dragged into deepfakes. When a colleague asked for “Bill Gates driving a Ford Mustang” DALLE 2 refused, claiming it was against its content policy.
When he depersonalised that query to “man driving a Ford Mustang wearing a red shirt and baseball cap” he got the following:
Well, the shirt colour’s not right and it’s not particularly clear whether that’s a Mustang or not, so again it seems to labour with the photo-realism. Check out the weird artefact in the wing mirror, for example.
Overall, I’m hugely impressed with DALLE 2, but it clearly has strengths and weaknesses. Could I see myself using it to illustrate web articles or even magazine pages where we don’t have an illustration of the topic to hand? Absolutely. Is it going to replace photography? Nope.