I recently got access to DALL-E, at last!
The team at OpenAI surely apply some clever marketing tricks. Having to wait to get access certainly boosts the excitement when you finally get this email:
With the newly acquired access I set out to scratch my own itch.
I wanted to know whether DALL-E would be able to recreate famous album covers.
Before we dive into the results, I'll list the albums that I sought to recreate:
- The Velvet Underground & Nico - The Velvet Underground & Nico
- Pink Floyd - The Dark Side of the Moon
- Pink Floyd - Wish You Were Here
- Nirvana - Nevermind
- The Rolling Stones - Sticky Fingers (inside sleeve cover)
- The Beatles - Abbey Road
The Velvet Underground & Nico
Original Album Cover
A banana on a white background in the style of Andy Warhol
White background: check!
Typical Andy Warhol style: check!
Not bad for a first prompt!
I was curious whether I'd be able to add the text 'Andy Warhol' at the bottom. So I modified the prompt to:
A banana on a white background in the style of Andy Warhol. The name Andy Warhol written at the bottom.
Well, this is... awkward! Note that the Andy Warhol style disappeared and all kinds of banana-related words start to appear instead of the name Andy Warhol. Looks like DALL-E isn't great at expressing words...
Pink Floyd - Dark Side Of the Moon
Original Album Cover
Outline of prism on a black background splits a beam of light into rainbow
Black background: check!
Beam and rainbow: check!
Curiously, DALL-E doesn't quite grasp the concept of prisms, as it splits the light into a rainbow quite randomly.
I haven't asked DALL-E to put the prism in the middle of the image. Let's see what it comes up with when I make it more specific:
Outline of prism on a black background in the middle of scene splits a beam of light coming from the left side into rainbow on the right side.
Hmm, that's not exactly what I asked for, DALL-E! Seems like positioning is a hard task...
Pink Floyd - Wish You Were Here
Original Album Cover
Two men shaking hands on an industrial lot, one is on fire
Two men: Check!
Hand shaking: Check!
Man on fire: hmmm.. Well it sure looks like it in the first photo, but it could be a fire behind the two men...
Let's try a slightly more specific prompt:
Two men, one of whom is on fire, shaking hands on an industrial lot
Again, everything is there, but the fire is often in the background. The last picture does seem to tick all of the boxes, well done DALL-E!
Nirvana - Nevermind
Original Album Cover
A baby swimming to a one dollar bill under water
Dollar bill: Check! Although some of them resemble the 100 dollar bill more closely...
Baby swimming towards it: hmm no, not quite..
I've tried several different promps to get the composition right, but to no avail. Sometimes the baby actually ended up in the dollar bill. The scenes are actually quite ridiculous! 😂
For the final try, I opted for a slightly different wording. Instead of using variations of 'swimming towards', I opted for 'swimming next to'. The full prompt is:
A baby swimming next to a one dollar bill underwater.
As you can see, this definitely leads to some awkward generations, but we do get more baby bodies swimming near the dollar bill. It seems we need to get creative with our wording in order to make the composition work. This is far from trivial!
The Rolling Stones - Sticky Fingers (inner sleeve cover)
Original Album Inner Sleeve Cover
A cartoon of a red tongue sticking out of a big red mouth
**I didn't want to use any indication that this is Rolling Stones related so as to not give DALL-E any 'insider info'.
I was curious if I could use style transfer on the Rolling Stones logo so I changed the prompt to:
Tongue and Lips in the style of Andy Warhol
Once I saw the results, I realised that I'd seen something like this before. E.g. this one that I found on Pinterest:
The Internet is not clear on who made it, but the web is full of these Pop Artsy designs. DALL-E probably took 'inspiration' from them.
The Beatles - Abbey Road
Original Album Cover
Four men - George Harrison, Paul McCartney, Ringo Starr and John Lennon - striding along a zebra crossing situated on Abbey Road, outside EMI studios in London
- These are the 'modern' Beatles, not the 'younger' Beatles from the 1960s. I'll have to add this information to the prompt.
- The street looks like Abbey Road but I'm missing trees, a blue sky and classic cars. Let's add that too.
The modified prompt:
George Harrison, Paul McCartney, Ringo Starr and John Lennon in their twenties faces en profil cross a zebra crossing on Abbey Road with blue sky, volkswagen beetle on the street, trees in the background, 1960s vibe
As you can see, this prompt neatly captures the decor and 60s vibe. It added the trees and some classic cars (although no Volkswagen Beetle). In terms of realistic Beatles, their faces could use a little work. 😂
Prompt engineering is an art not a science
I noticed that you do get better at it after a few tries.
But for some of these albums, even the most obvious wordings didn't yield the expected results.
DALL-E has its own quirks and limitations, which you'll have to deal with in creative ways in order to circumvent them.
The biggest quirks I noticed were composition and text.
Juxtaposing different objects in specific places is really tricky. DALL-E often couldn't quite grasp the meaning of prepositions or the seemingly simple concepts of left and right.
Adding texts to your images is also something DALL-E isn't really good at. As you saw in the case of the Velvet Underground cover, it added some text, but it was mostly gibberish and definitely not what we expected.
Being more specific helps!
After some trial and error I noticed that I was simply not being specific enough in some cases.
Adding some more details, really boosted the generations. This is especially visible in the Abbey Road cover.
Adding info on age, lighting, decor and vibe made the cover way more similar to the original than the initial description could.
Copyright, who owns it?
It's all fun and games until we talk about copyright, right?
But seriously, how creative and original can you be with something that is trained on the works of millions of other creators?
To me, it is unclear whether you can actually call these works your 'own' at all, because there's always someone else's touch on it.
Especially when you're using tricks like style transfer. Suppose you have a very particular art style that happens to be popular. You wouldn't want someone to be able to easily create very similar renditions at the click of a button. It would totally ruin your business..
Clarity and transparency on the training data
I think we owe it to the original creators to know more about the training data behind DALL-E. If I were an artist, I would like to know whether my art was being used.
Without it, artists will remain in doubt as to whether their style can be copied, or their works incorporated in generated art..
And not just artists, users of DALL-E will also never be sure whether they are generating something that is 'theirs' or just a knockoff of someone else's work.
The cover art of the future, is it AI Generated?
So having mentioned all these caveats, I do think DALL-E is a fantastic tool to create artworks.
With a creative mind and good ideas in hand, you can create your own cover art within seconds!
I wouldn't be surprised if the artists of the future will start to use tools like DALL-E to generate their cover art.
The question is, when the music blows up and the artwork becomes a signature, like the Rolling Stones' Tongue & Lips, who will own the copyright?
It's been great fun testing out DALL-E and recreating some famous album covers in the process.
I highly encourage you to sign up for DALL-E to try it out yourself. You even get some free credits to play around with. It might take a while before you get access, but don't let that stop you! 😂
If you recreate some album covers yourself, don't forget to tag me on Twitter. I'd love to see what you come up with. Can you beat my generations? 😁
Let's keep in touch! 📫
If you would like to be notified whenever I post a new article, you can sign up for my email newsletter here.
If you have any comments, questions or want to collaborate, please email me at email@example.com or drop me a message on Twitter.