Using DiffusionBee on my Mac, I entered a photo of myself using the Image to Image feature, then entered the prompt Dressed as clown.
It produced this nightmare-inducing result:
Original photo as reference:
I got a fancy cap in the process, along with a dress shirt (?), and it made the background more overgrown.
I will experiment more, but maybe with fewer clown prompts. Or more. I haven’t decided yet.
Not really, I’m just being dramatic.
In Stable Diffusion, I entered this prompt and set it to 50 steps (the default is 25):
sunrise over a verdant field in the countryside. Oil on canvas.
And I got this lovely pastoral scene about a minute later:
And I thought, “How long would it take me to actually paint this, even using a digital canvas?” It would not take about a minute. It would take many minutes. It would take hours. It would take time to experiment with colour and tone, adjusting texture and light. But I feed the text into the AI and it produces something that is actually pretty good–probably better than I could do (I was never great at using paint as a medium) and it gives me pause and think about all the articles out there raising flags over AI-generated art.
I must ponder this. I currently think of AI-generated art as an actual form of art, but a different one. It requires a human hand to do anything (ostensibly, at least) and the prompts can result in dramatically different results with some seemingly minor tweaking. There is art in the crafting of the prompts themselves. But still, I get why people are concerned. This stuff is already shockingly good, and we are in the early days.
Does this mean some guy will no longer have to paint hundreds of black velvet Elvises to make a living? That it can be done in moments by an AI instead, freeing him up to do something more meaningful, possibly with orphans? Or does it just mean this guy has lost a job?
Ponder, ponder.
(And yes, I will do a black velvet painting of Elvis prompt later.)
(Also, amusingly and fittingly, the spelling checker did not recognize “Elvises” and suggested “pelvises.”)
I used this Github one-click installer to get Stable Diffusion, another AI tool that generates images based on text prompts, running on my M1-based Mac.
The initial results have been amusing, interesting, and sometimes just weird.
Using the default parameters, I used this prompt:
Cat wearing a top hat
And got this:
I then used the same prompt, but set the steps from 25 to the max of 50:
This is one dapper feline.
Next, I tried:
Dog on a cow. High detail, photorealistic.
And got:
As the kids say, there is a lot to unpack here.
Undeterred, I adjusted the prompt:
Small dog riding on top of a cow through a grassy field.
The result looks like an alternate version of the album cover of Atom Heart Mother. I like it.
I’m going to keep playing with this, since my invitation to join the DALL-E beta went into a black hole, never to be seen again. More images soon™!