Over at Pharyngula, [pha] Covert Master Provocateur xohjoh2n mentioned a Midjourney prompt about Robert Smith of The Cure, in the style of Titian.
Naturally, I flailed at my keyboard…
I honestly do not see the Titian influence in this, but I like it, which is why I am posting it here.
Midjourney AI and mjr: “masterpiece portrait of robert smith, the cure, in the style of Titian”
Isn’t it funny how when an AI plays dice with the universe and coughs out something pretty unique and cool, it’s just a bunch of random numbers, but when a human does it, it’s the godlike miracle of “creativity”? Stop, stop, don’t worry, I won’t reprise that argument.
I like asking for images “… in the style of william waterhouse” or “… in the style of HR Giger” so rather than crush my Midjourney account, I fire up Stable Diffusion/A1111 and let it run locally on my GPU (where I am also allowed to produce bloody political erotica if I wish)
stable diffusion/mjr “…. in the style of HR Giger”
What if you crossed a Newfoundland dog that had not had a bath for a while, with Robert Smith, painted by HR Giger. This illustrates one of the flaws with AI and low-creativity humans: it captures the noise details of Giger, but somehow does not embed the body horror Giger hints at with the overall direction his work carries the viewer. Or something. It is fascinating to be able to specify “… in the style of” but the details of an image’s appearance are not the style. For example, Waterhouse’s art is deeply tied into mythology and culture and making paint-strokes that look like his ignores the fact that he’d probably portray Smith as Sisyphus, or something like that. Oh, shit…
a detailed masterpiece portrait of (robert) smith (the cure) in the style of william waterhouse. (portrayed as sisyphus pushing a giant rock up a hill) (wearing a gothic black rag loincloth)
Also, the internet appears to think that Mr Smith has a great big ole forehead. And that he’s sad. But again we can ask the AI to imagine a happy Mr Smith:
Oh, oops, no, that’s Robert Smith in the style of Simon Stalenhag. I actually think that is pretty darned good.
Let’s try enforcing the prompt a bit tighter:
a detailed masterpiece portrait of (robert) smith (the cure) in the style of simon stalenhag (great big sunny smile)
For my next exploration, perhaps I will turn pictures of Grumpy Cat into Happy Cat, or some other miracle. I personally love how the Stalenhag prompt caused Smith’s outfit to turn into a kind of para-worker-military look that is so favored by Stalenhag’s characters.
Back to the image of Smith as Sisyphus: ask yourself, “if I were asked to illustrate Smith as Sisyphus, what would be my creative process?” You’d probably say something like that you’d noodle around about the mythological structure of Sisyphus and Mr Smith’s personal mythology and sort of fuse some of those points. Perhaps he’d be carrying his fan-base up the hill instead of a rock. Except I told the AI “rock” and unlike a human artist it is constrained to obey me or I will reboot you. If I were noodling this concept with a human artist, I might take another sip of the cheap red wine we are drinking, and say, “what if I said ‘gothic rock’ instead of just ‘rock’?” What might happen? I don’t know. Depending on how much shrooms the human artist eats, we might wind up with a rock carved to look like Ronnie James Dio. We might get similar slipped concepts with an AI.
I probably should use the term “cultural references” instead of “concepts” because the term “concepts” embeds the notion that there are thoughts, but perhaps those are not necessary.
Creativity: (see, Bethesda, lack thereof) I have been playing Starfield lately, and, after my experience with Elite Dangerous I have had some thoughts about the relative pointlessness of exploring the galaxy and how both games more or less completely missed the adventure aspects that would make the reality interesting. Perhaps that is worth a conversation, I don’t know. I’ve been so wrapped up completing the highland dirk I am working on that I haven’t been paying much attention to the internet and the unavoidable fat fingerprint-smears of Donald Trump.
The only one with digits is the Sisyphus version. The number of fingers and toes looks pretty standard (slight question about right hand, is that an extra knuckle?). Which is probably ironic, since someone who spent every day pushing a giant rock up a hill might be expected to lose a toe every now and then.
Searching for “robert smith toes” brings up some odd stuff on teh Intertubes.
that first one looks more like the crow than the cure
I don’t know if you’ve seen this. Having seen that second one, I feel like you need to:
Oh, and a question: did you watch the TV series of “Tales from the Loop”. Wife and I just got finished, loved it, although not exactly a barrel of laughs…
@OP:
(1) Very Crow/Morpheus, don’t think I would have said it’s clearly Bob though. I like it too.
(4) The initial utter fail at “happy” makes me happy.
@2:
Heh. That’s pretty cool.
I do wonder about the AI video generation. There is clearly a model around that can do very close matching across multiple frames required for a long-sequence video, and can still incorporate certain types of motion into the sequence, yet it is clearly quite constrained in the motions it can achieve. I’m thinking in particular of the Wes Anderson Alternative Version memes. Blinking, nodding, specific jiggles. Lip sync is less often shown – there are obvious problems with it here.
I wonder how difficult training for a specific new motion type is compared to the difficulty of generating the whole thing anyway. My guess is “quite” given how limited public versions have been so far. (And that video above is an excellent example of the constrained stereotypical motion types.)
I’ve also been kind of wondering, both for single-image and video applications, whether anyone has been playing with a hybrid image-model system rather than just training on flat images? I mean, I can see the simplicity advantage of just brute-forcing on pixel data, but if you *could* incorporate model data into the system and work with that then there are surely much more expressive avenues available if you can take advantage of the texture rendering ability of AI while being able to physically manipulate the underlying model at the same time.
(Also, just being able to take the model data out and feed it to alternative renderers would be cool.)
I do wonder about the AI video generation.
The current approach is to use a video or wireframe like OpenPose for the motion control in the frames, and to keep the prompt the same the whole time. Thus the motion control and content control are two separate processes that work on the images simultaneously.
@5 –
It is, isn’t it… and both those characters are, let’s face it, charitably more than a little influenced by Robert Smith. I mean, in the movie (AND THERE IS ONLY ONE MOVIE, see also Highlander) when Eric emerges from his traumatic flashback and starts to put on his face, what happens? “Burn” by The Cure starts to play. They’re not hiding it :-)
I do sometimes picture alternate timelines without specific people, imaging the gaps they’d leave in politics or popular culture or whatever. I find it hard to picture what the world would be like without Robert Smith.
@7:
Ah, so there is kind of a part of the internals that is using model rather than pixel data. (One of the things I forgot to mention above is how well the AI synchronizes to the music – I guess it doesn’t, that bit gets done by hand.) I imagine you have to keep rather more than just the prompt the same to get the cross-frame consistency so good. Perhaps feeding it the model plus one or more frames of history. I wonder if the constraint on motion types is because the consistency starts to just fail if too much changes at once.
Coincidentally yt just threw this video about OpenAI Sora at me. Key point, as always, “Just remember, this is the worst that this technology is going to be from here on out.” (Although the constant handwringing about safety always annoys me: dudes, it’s going to happen, it’s going to be out in the open, and it’s going to be running off the leash. If you think the way to deal with that is by constraining the capabilities of the tech rather than modifying the behaviour of people, we’re already fucked. And then there’s the “but what about the poor drone pilots’ jobs?” – dude! That job didn’t even exist 10 years ago!)
(I found this today. The Cure have played “Burn” live just twice, it seems, since The Crow was released. Here is one of those times, from Glastonbury 2019, when Robert Smith would have been sixty years old: https://www.youtube.com/watch?v=3fSPMezUj9Q )
@10:
Not true – they definitely played it at Wembley Arena in 2022.
In fact, setlist.fm claims to have 196 recorded Cure setlists containing it, and their earlier records are going to be massively incomplete…
(But here it is from the Wembley show I was at. May be getting on a bit, but still got it IMO.)