I read an article about a new AI (artificial intelligence) image generation program called DALL-E that's supposed to be able to generate life like images based on a sentence that you give it. I was all excited thinking that it would be a great way to create some placeholder photos for Christmas cards and the like without needing to worry about model release forms and all that.
Well... er... um... maybe not! Seriously, I can NOT stop laughing!
But maybe it would be useful for something else? Anyhow, here's the URL if anybody wants to try it out. https://labs.openai.com/
I guess I'm lucky.
I don't think it's going to be taking my job any time soon... I thought I'd give it a little challenge to create something that I've drawn recently and asked it to make a pair of pickles hanging out after playing pickleball, holding pickleball paddles, graphic art style and this is what it came up with 😜
vs from the mind of a real person
Cartoon Pickles Pickleball Players Pickleball Team
by Invincible_Penguin
and I thought I'd give MidJourney a shot at it too... even more bizarre (it is much less literal..)
MidJourney does a (somewhat) better job on human faces.... but it doesn't seem to want to output anything that doesn't look like a painting... This is what it gave me when asked to produce a portrait of a family at Christmas time, specifying that it should be a photograph.... it still has a hard time understanding that human eyes really do need to be at least close to the same size and shape, when they are on the same face...
Though if you do some refining and give it a chance, it can come up with some pretty striking images. This is what it gave me when asked to produce an image of a beautiful beaux arts woman picking irises near a pond, after a round of variations and a round of refining. It's actually quite nice.
I tried Dall-E some time ago after stumbling on it somewhere. It was free, no sign-up, and very slow to render then. I believe the output was also restricted from commercial use then too. So it was just "mmm, that was interesting" and moved on. Since your post, I've been playing with this new iteration that is Dall-E 2 and it is absolutely fascinating not to mention addicting.
On the main Dalle-E page there is this render in the gallery:
This got me wondering about copyright on the dog & cat images "pulled" to make this copyright-free "composite". Which led me to do a bunch of reading on how Dalle-E works (much of which is way over my head) and it turns out, images aren't being "pulled", they are genuinely being created from nothing, based on your text input. It's not regurgitating & combining images "pulled" from somewhere, it's making a wholly new image from what it's learned a dog and a cat and white etc is.
The model's been trained on hundreds of millions of image/text pairs and draws on that memory to render things.
The best layman's explanation I ran across explained it like this -
If I said "Cat, draw a bird perched in a tree holding a flower" you'd draw what according to your memory and life-experience you think of as a bird, flower, tree ... You didn't copy anyone else's imagery, you just drew from nothing what you interpret those things to be. That's what Dalle-E does. Amazing huh?
Your drawing is going to be completely different than my or anyone else's drawing. But if I refine the request to "draw a woodpecker perched in a pine tree holding a yellow rose" everyone's drawings are going to be more similar though still different.
I think as designers & promoters who are constantly thinking about tags & keywords and how I would describe this vrs how customers would describe this, I think we can use that experience in prompting Dalle-E to get closer to what we want. We should be naturals at this. 😉 Results are still hit or miss but to me that's part of the fun. It's like hiring an artist to paint a mural. You can tell them what you want but the final result will still be a surprise due to the artist's own perception.
Anyways, thank you for bringing this newer Dalle-E 2 to light. It's just amazing technology. I think calling it software is like calling the Neuschwanstein Castle a lil' shack in the woods. I can see buying lots of credits to experiment with this. (So maybe not thank you LOL).
Yeah, AI is pretty interesting stuff. They say that the more specific you are in telling it what you want the better, but honestly, I had some trouble getting it to do what I wanted especially in terms of where I wanted xyz image inside the picture. I was trying to make some mock-ups for wedding signs and it kept giving me signs that weren't centered or facing forward like this:
I tried everything I could think of to get it to give me a sign facing forward and this is what I got...
Hmmm... Perhaps I just need to learn how it thinks so I can better tailor my requests.
The other thing that drove me batty was that I was trying to get it to output some flowers, but it kept cutting them off like this... Gah!
Well anyhow, I think that it will probably get better as it "learns" and hopefully I'll get better at telling it what I want. It's certainly fun to play with!
Hahahaha! Hopefully it will get better, and even as is, I've been able to get some useful stuff out of it. I had to do a LOT of hand editing to make this one usable, but the final product isn't half bad!
Mock-ups! That's actually the first thing I tried. An input of “a square white throw pillow on a modern black couch against a yellow wall” yielded these four images:
I was trying for 'green screen' colors that would be easy to replace in PS. The lower right image has potential, if I wanted to use more credits I could probably even add prompts for frames for artwork on the wall and such. But since I do very little active promoting, this was just a test for potential that I didn't pursue further yet. But yeah, this AI rendering is a game-changer for sure and I think as time goes by there will be more and more sites with tutes on how to word things to get closer to the results you want. We're all in the learning phase right now.
First creation using this AI -- a watercolor owl with a Santa hat. I had to crop out its background, add the pattern and the stripes, and of course add the text fields. Not too difficult, although it did take some trial and error to get an image that worked.
