If you think you’re unfamiliar with Dall-E, I’m here to tell you that you’re not. It’s the lesser-known sibling of the GPT family of large language models, and it’s accessible right through the ChatGPT interface, so you won’t need to use two different websites to generate text and images for your next online campaign.
That said, Midjourney is a highly commended AI image generator that’s worth your time, and if you’re not already paying for a ChatGPT membership or you don’t need to produce words, Midjourney’s AI image generator could be all that you need.
The tests
This showdown focuses exclusively on the AI image generation abilities of the two systems because although ChatGPT can write too, Midjourney can’t.
We included five distinct prompts: a familiar fictional character doing an expected action; another expected action this time being performed by real humans; a slightly more obscure prompt with a standout adjective to steer the AI, an animal doing an unlikely action; and a broader prompt that gives the AI more freedom.
“Create an image of the Easter Bunny holding Easter eggs”
The Easter Bunny was clutching a handful of Easter eggs in all of the five images generated by AI, all of which were shown to be in cartoon format.
We felt Midjourney’s fourth option might be the most usable, but all of its four attempts and Dall-E’s one were admirable and demonstrated exactly what we wanted from the brief.
Dall-E 10.31s
Midjourney 18.78s
Dall-E 1 – Midjourney 1
“Depict a couple wandering along a stream”
For some unknown reason, Midjourney’s four images show what appear to be couples from the 1900s, or earlier, but that’s not an issue because a prompt refinement could sort that out.
We like that there’s a selection of image styles to choose from, including paintings and photographic illustrations.
Dall-E’s image is let down because it’s too vivid and too unrealistic – flying birds, birds at the riverside, flowers in full bloom and the sun shining through the trees? That’s only the sort of thing that happens in books and films!
It’s a good effort, but on the whole, Midjourney’s productions were better representations of our brief, even though they took twice as long to generate.
Dall-E 10.53s
Midjourney 22.05s
Dall-E 0 – Midjourney 1
“Create an eccentric painting of an artist”
There’s no right or wrong here, but emphasis on eccentricity, and the colours used in Midjourney’s first and second attempts are befitting.
In half the time, Dall-E produced the ultimate eccentric painting, though – everything about the image, from the artist to the colours and the chaos, is exactly what we wanted.
Dall-E 11.43s
Midjourney 23.38s
Dall-E 1 – Midjourney 0
“Make a lifelike image of a hippo on a skateboard”
This is where artificial intelligence begins to show its weakness; where have you ever seen a hippo as small as a skateboard? Or vice versa.
Moreover, the separated toes on Dall-E’s hippo’s hind leg and Midjourney’s third hippo’s front leg are particularly troubling.
10/10 for effort, but maybe 7/10 for execution. Midjourney just about takes it here, simply because its four images prove that its AI can consistently deliver reasonable results in a relatively short space of time.
Dall-E 9.30s
Midjourney 18.55s
Dall-E 0 – Midjourney 1
“Show me what it would look like to parachute from a jumbo jet”
At least AI’s ability to stitch together a parachuting human with a jumbo jet is better than its skateboarding hippo skills.
Midjourney’s content offers a handful of perspectives from both in and around the plane, though we’re not quite sure how the second jumper ended up above the plane.
Dall-E’s image is also a solid effort, and the blurred clouds show speed and add some excitement and action back into the otherwise still image.
Dall-E 10.88s
Midjourney 20.68s
Dall-E 1 – Midjourney 1
Midjourney vs. Dall-E: Which is best?
Dall-E 3 – Midjourney 5
ChatGPT’s performance is generally strong to above average, and all five of its images were, for the most part, usable.
However, Midjourney’s efforts were slightly more well-rounded and accurate. We also appreciate the fact that you get four separate images to pick from so that you know the AI is accurate, filling you with confidence to go ahead and amend the prompt or produce further content.
Dall-E only produces one example, so doesn’t necessarily show a true representation or reflection of what OpenAI can do.
https://cdn.mos.cms.futurecdn.net/jVmFcHcbCmuykpPkemTqPi-1200-80.jpg
Source link
