Comment by vova_hn2
1 day ago
I find it kinda funny that even "smartest", most capable models often have a very hard time going beyond surface-level, top-of-mind associations, when faced with creative tasks.
Of course, a "Japanese Minimal Poster" has a sakura and a stylized flag of Japan, duh
A human designer would probably think that "Japan -> sakura" and "Japan -> Japanese flag" are way too obvious, too banal, too stereotypical, and, most importantly too boring. And then they would probably sit for a while and try to think of something fresher and less clichéd.
But AI has absolutely no problem with going with the cheesiest, most overused trope.
Author here.
Well of course Spring + Japan = Sakura. In a way the model would be faulty if it didn't reach for that cliché. That's what it's meant to do.
If you want originality, you need to supply it. AI fills the vacuum with the obvious thing.
And in a way that's a feature. If I was creating a poster myself I'd think "nah, can't use sakura it's too obvious". AI gets you past that hurdle. "Yeah, we used the cliché, it looks nice, let's move on".
This is if it's for a school fayre, village fête, etc. - brilliant originality isn't the goal. But it's nice when your poster doesn't look the same as three other posters on the same noticeboard (I have seen this).
If brilliant originality is needed, a human has to bring it. Although an LLM with the temperature turned up can offer some nice seeds.
If you need pseudo-originality I have a skill for that: https://github.com/iandanforth/force-associations
Nice! I use an old madlibs library I built for D&D: https://github.com/gavmor/Dunmanifestin
2 replies →
Tyty I was thinking about something similar
Association clouds.. Routes.. Whatever this is..
It is a prime example of esoteric LLM incantations, if not -excuse me- nonsense
1 reply →
I love it. Entropic prompt injection
"AI fills the vacuum with the obvious thing."
I like that a lot.
It literally is how they work. They look for the next most probable token. Always keep that in mind when writing a prompt or reading output. If you don't supply it, it will fill it in with the most obvious stuff.
Personally I think if you were going to make a poster for a Japanese themed spring festival, and you did NOT include some sakura, you completely misunderstood the assignment.
Your promotional poster for a community gathering should not be subverting the western gaze on Japanese culture, or any other concept from art criticism 101.
It's supposed to get average people in your neighborhood to glance at it for 5-10 seconds and say, "cool looks like a good time"
YMMV but my experience tells me that the more cliché something looks, the more likely it's bad/cheap.
I've never been in a cliché-looking restaurant that is actually decent.
Meanwhile, the best restaurants I know have, at best, a flag or some kind of small drawing to give a clue about their nationality, but otherwise they look like a regular local one.
> I've never been in a cliché-looking restaurant that is actually decent
You either have an absurd standard for "decent" or this is one of the weirdest comments I've ever read here.
Right, but the best restaurants are not most restaurants. The best high end events are not most events. Most events and restaurants could be improved a lot by AI already
Artist here. I'd rip off Kawase Hasui 100 different ways without using sakura or subverting anything.
More like 100ms or less as their visual field coincidentally sweeps over wherever the ad is posted in the physical world. In this regard, it's even more important to use something people subconsciously associate with your thing.
That's a prompt problem.
If you named a Japanese artist or style rather than giving an adjective derived from a country, you might get something that wasn't national wank.
Yes, a human might do this step for you. A mediocre one might not.
> Yes, a human might do this step for you
An artist (even a mediocre one) almost certainly would do this step for you
Yeah because... AI is a tool, not an artist. In other words water is wet.
2 replies →
Ah yes they should have gone with that one wave
Isn’t the most overused trope the most statistically weighted to be chosen from a group of options? It makes technical sense but as you say leads to boring outcomes
If given an underspecified option, yes. Which is why it's good to specify options.
I see what you mean, but I think you can draw (no pun intended) a distinction between trope following and having a skill, or at least should notionally be able to.
> But AI has absolutely no problem with going with the cheesiest, most overused trope.
Considering your prompt fits exactly that description I think the AI delivered perfectly. Try using some actual design language and see how well it does.
The thing is, they aren't creative. When I ask an LLM to go beyond the mean, it has no idea what to do and returns incoherance. Its responses seem to have N latent dimensions of information. But, the value of each of those dimensions is always the midpoint. With creative prompting, I can push it to respond in ways that go beyond the mean, but it can't tie them together cohesively. It doesn't have anything in its training data to fill the hole in space. No ability to extrapolate.
My point is that those kinds of prompts are dumb and I actually expect the model to produce slop. In programming we write these huge prompts that go into a lot of detail, and when we do that, it actually works. I've never asked a coding agent to go beyond the mean, that just sounds like a dumb thing to do. Apply this rule in design and I'm sure you'll get better results. Kinda what the blog post is all about.
> But AI has absolutely no problem with going with the cheesiest, most overused trope
I genuinely wonder if this can be addressed (at least partially) as simply as prompting: "don't rely on overused tropes", or something to that effect.
The masses love clichés and stereotypes, which is why these "artiste" designers frequently miss their target audience.
That's on the prompt, not the AI. You could be specific about what imagery to use like a torii or a Japanese castle, and so on. Whatever you leave to the AI you're most likely going to get the most generic version. If you're low effort in the prompt yet get low effort results, if you put some effort into the prompt you can get better results.
If you understand how curve fitting works, then it is a very predictable outcome.
Sure, but what's interesting is if you tell it to "behave unpredictably" or "don't use tropes" it usually fails down the same curve. Which is predictable, but telling.
I remember seeing a poster for Chinese New Year years ago, and someone had included the kanji for Japan on it (日本). Obviously they just wanted generic looking characters without bothering to see what they meant. I couldn't resist writing a comment on the poster about it.
It showed whoever did it had no clue about East Asia or Chinese script...
Are you insinuating those 1970s fonts takeout fonts aren’t actually how Chinese people communicate? I knew they were AI!
for the price you pay ai you would 100% get a cherry blossom tree from an inexperienced designer
Ai costs under $10 though, you wouldn’t be able to hire a human for that
Pretty sure 90% of the posters of the type at the top of the article are made for $0. I bet for $10 you could get real artists from non-Western countries to make a simple poster, but these days people would probably claim those $10 jobs and then have AI generate them anyway.
right, even worse... if you paid 5X that you'd still be getting whatever was the first idea that came to mind
You're easily in 10X territory to get something pretty basic with a round of feedback and maybe 50X to start to get something good
the low to mid range of any visual design market has essentially been destroyed
1 reply →