That GAN really tied the room together, man.

September 2, 2021 9:19 AM

AI Movie Posters : Can you guess the movie from a Dali-esque image generated from a text description of the movie?
posted by kaibutsu (58 comments total)

I managed 13, I think.
posted by pipeski at 9:28 AM on September 2, 2021

I would really love to see the descriptions fed to the AI to generate these, because a lot of them are almost too damn good. The ones that were obvious to me were obvious at absolute first glance, like looking at the movie from a passing train.
posted by uncleozzy at 9:32 AM on September 2, 2021 [20 favorites]

I got all of 5. And it wasn't the ones I would expect. I thought Die Hard was either The Dark Knight Rises or Wolf of Wallstreet. Monty Python and the Holy Grail made me think I'd missed a Redwall movie. I figured Rear Window for either Dred or The Raid.

But I got the Hurt Locker without ever having seen it. Ditto National Treasure and Point Break.
posted by Hactar at 9:34 AM on September 2, 2021 [1 favorite]

Being John Malkovich was the funniest - and easiest - one.
posted by dragstroke at 9:35 AM on September 2, 2021 [6 favorites]

Interesting how the Dr. Strangelove poster was in colour.
posted by cardboard at 9:52 AM on September 2, 2021 [2 favorites]

I got 13, but mostly what I realized was that I either got it instantly or not at all. Wasn't really any middle ground.
posted by BlahLaLa at 10:02 AM on September 2, 2021 [9 favorites]

I got 5 or 6, and there were about the same number again that seemed instantly obvious in retrospect. Can't believe I missed Space Jam, for example.
posted by Paul Slade at 10:07 AM on September 2, 2021

Only got four, only Big Lebowski ans Space Jam seemed really obvious. Hilarious that Cast Away nailed the FedEx color scheme.
posted by bendybendy at 10:08 AM on September 2, 2021 [2 favorites]

So GAN isn't able to create a very convincing representation of any thing in particular, but I find it incredible that it's able to paint a good enough impression of the words that people can guess what it was thinking. It's like playing pictionary with the child version of the future overmind
posted by Popular Ethics at 10:08 AM on September 2, 2021 [7 favorites]

I got twenty—which I’m amazed by, considering how bonkers the images are. They get easier as you go along, I think. Fun!
posted by EXISTENZ IS PAUSED at 10:13 AM on September 2, 2021

I guessed Fear and Loathing in Las Vegas for all of them, but only got one right.
posted by aubilenon at 10:13 AM on September 2, 2021 [26 favorites]

I'm also curious about the text that was fed into the GAN. I'm guessing it might be the Wikipedia synopsis or something along those lines? The "Back to the Future" one includes a clock-tower, which is a big part of the film but I don't think one that I'd include in a one-sentence summary. "Ferris Beuller" has Chicago Cubs logos everywhere, which, okay, it's set in Chicago, that's a reasonable inferrence. "Miami Vice" has some neon pink accents that I associate with contemporary-things-trying-to-look-like-the-80s. I wonder what text got the GAN there.
posted by Alterscape at 10:20 AM on September 2, 2021 [2 favorites]

I got 12, and oddly some of the ones I got I haven't actually seen. Like Mamma Mia. Never saw the movie, but that's immediately what popped in my head when I saw the picture.

A few others I didn't get, but was pretty sure of genre or setting (like heist for Oceans Eleven or Chicago for Ferris Beuller).
posted by ghost phoneme at 10:25 AM on September 2, 2021 [2 favorites]

The color schemes in several fit too well for it not to have access to that information in some way or another.

I'm amazed you all are doing so well, these all look like visions from a perpetual drug-induced psychosis to me. So yeah, Fear and Loathing in Las Vegas.
posted by skewed at 10:27 AM on September 2, 2021

Oddly enough, I got Sideways straight away.
posted by pipeski at 10:28 AM on September 2, 2021

Wizard of Oz.
No Tin Man, No Lion and what seems to be an either an AR-15 or Bissell brush boss in Dorothy's hand
posted by clavdivs at 10:30 AM on September 2, 2021 [2 favorites]

> I wonder what text got the GAN there.

One thing to keep in mind is that these models aren't pristine ones being trained from the ground up. They were originally trained on (most often in these newer ones) text and images from the internet. (For generated texts we see online "trained on The Office scripts!" or whatnot, the training done to produce the various things is more fine-tuning the original model than direct training.)

So a lot of the associations are already there, and if a chunk of text that's fed in to generate the images manages to match up in probability with stuff it was previously trained on (that is a garbled and oversimplified description of what's going on), that's what gets the GAN there.

Coincidentally, this morning I've been fine-tuning a GPT-2 on a dataset that's a combo of Warhammer 40K and shifter romance book descriptions, because I am just that edgy. It has twigged that I'm describing science fiction and fantasy franchises and the sample stuff it's thrown out so far has drawn from Harry Potter, Star Wars and Discworld. (And once, nail polish.)
posted by telophase at 10:34 AM on September 2, 2021 [3 favorites]

I thought I couldn’t, but then I tried reorienting my brain and got a few. I like the first one for Spirited Away, though of course that isn’t correct.
posted by Going To Maine at 10:37 AM on September 2, 2021 [2 favorites]

Wasn't really any middle ground.

I got a suprising amount right, but I guessed Polar Express for the first, only to find out which was Snowpiercer, which...sure, that actually makes sense!
posted by Jon Mitchell at 10:52 AM on September 2, 2021 [3 favorites]

I thought Mamma Mia! was Titanic.

And that's when I quit guessing and just looked at the pretty pictures.
posted by chavenet at 11:09 AM on September 2, 2021 [8 favorites]

I got 8 and that freaks me out.
posted by mazola at 11:23 AM on September 2, 2021

Whether I can guess them or not, these are really cool.
posted by jacquilynne at 11:26 AM on September 2, 2021

but I guessed Polar Express for the first

You're not the only one. Calling it now: Snowpiercer is just a gritty reboot of Polar Express.
posted by Faint of Butt at 11:28 AM on September 2, 2021 [6 favorites]

I got Willy Wonka because of the purple, and Space Jam because of the basketball-ish things. The rest were completely unguessable for me.
posted by ardgedee at 11:52 AM on September 2, 2021

Snowpiercer is just a less terrifying reboot of Polar Express.
posted by All Out of Lulz at 11:55 AM on September 2, 2021 [8 favorites]

I was surprised at how well the “computer” did at generating these images, but I’m at a loss when it comes to rating the significance of the achievement. Is this a “wow!” or a “hunh, sure.” situation?
posted by Don.Kinsayder at 12:16 PM on September 2, 2021

Is there anyone else here who just saw a big jumble and couldn't come up with a single one?
Something must be "wrong" with my visual processing
posted by tigrrrlily at 12:22 PM on September 2, 2021 [2 favorites]

11. Thé one for Office Space is good enough to be the actual poster.
posted by CheeseDigestsAll at 12:28 PM on September 2, 2021 [2 favorites]

I love these, I would legitimately put them on my wall. Office Space is one of my favorites. I also guessed Polar Express for Snowpiercer, and Titanic for Mamma Mia! One of the most surprising for me was Star Wars, because it somehow looks absolutely nothing like Star Wars and yet is to me also very obviously Star Wars.
posted by biogeo at 12:36 PM on September 2, 2021 [2 favorites]

I think the Star Wars one must have gone pretty heavy on "Princess" Leia as an element.
posted by team lowkey at 12:37 PM on September 2, 2021 [6 favorites]

I thought Star Wars went heavy on "Star", as in movie star on the red carpet, or Dancing with the Stars.
posted by telophase at 12:40 PM on September 2, 2021 [1 favorite]

Also, seconding telophase's point that the model used to generate these probably already "knew" quite a bit about what these specific movies should look like: images and posters of the movies attached to descriptions probably formed a significant part of its training set, so in some sense the model is just being asked to reproduce its training data. I'm not terribly shocked by the fact that it can do this recognizably, but the very impressionistic way in which it renders the images is really cool.
posted by biogeo at 12:43 PM on September 2, 2021 [1 favorite]

team lowkey, you're right! I was really wondering what the pink skirt looking things were doing there, that's got to be it.
posted by biogeo at 12:44 PM on September 2, 2021

Briefly reading the accompanying description the special sauce here seems to be that they use one AI which is really good at stitching images together into a coherent piece, and then another AI that is really good at matching images to descriptions.

So the first AI is doing its Google Image Search on all of the words in the description (which, yes, is going to include a lot of results of images for the movie itself), tying them together, and then the second AI scores how good of a job it's done. So we're seeing the most successful combination of the two.

The first AI grabs all the things from "An F.B.I. Agent goes undercover to catch a gang of surfers who may be bank robbers." and the second AI says "Yes! That describes this picture!" or "Try again".
posted by team lowkey at 12:51 PM on September 2, 2021 [3 favorites]

telophase: "I thought Star Wars went heavy on "Star", as in movie star on the red carpet, or Dancing with the Stars."

It looks like a bunch of birthday candles to me, but I can't guess where that came from. It would be really interesting to see what descriptive text they used.
posted by team lowkey at 12:54 PM on September 2, 2021

I think team lowkey must be right about Star Wars -- it's hard to see where all the pink skirts came from if not from "princess".
posted by tavella at 1:05 PM on September 2, 2021

Call Me by Your Name just peaches making out on some terrace, very funny artificial humour
posted by Ten Cold Hot Dogs at 1:26 PM on September 2, 2021 [1 favorite]

From how instantly I got Willy Wonka, I'm guessing the ease of figuring out the movie is in proportion to the number of memes that use images from that movie.
posted by fnerg at 1:38 PM on September 2, 2021 [7 favorites]

I'm not even going to start trying, but these are all incredibly beautiful. (Edit: oh, got Hurt Locker immediately, and of course Wonka)
posted by turbid dahlia at 2:39 PM on September 2, 2021 [2 favorites]

It’s like reverse surrealist Dixit!

I think I got about four. But those were nearly as obvious to me if the title were on the poster. There’s definitely a brain clicking thing going on, like recognising a face. The ones I didn’t get were completely unparse-able and confused my brain, like a cross between magic eye and abstract art.
posted by iamkimiam at 2:39 PM on September 2, 2021 [1 favorite]

I got Space Jam and Willy Wonka but that was after I gave up and went and read the thread, then went back and tried again—so I knew to expect them.

The rest I couldn’t even begin to fathom. I don’t watch many movies, which is maybe part of it, and I’m bad at recalling the names of things, and I have the “processes images as a collection of details rather than seeing the whole picture” autism thing. Would be really interested if other autistic people were similarly baffled.
posted by brook horse at 4:32 PM on September 2, 2021 [1 favorite]

To clarify: they mostly all made total sense when I clicked on the title. But I couldn’t even organize any of them into coherent shapes before that. I couldn’t tell that the Oz poster had a girl and a house, for example.
posted by brook horse at 4:33 PM on September 2, 2021 [1 favorite]

I only got a few but a vast majority clicked fire me once I hit the reveal button.

It reminds me of how, in dreams, you can recognize something wonky as a person or a place or a thing but it'll only strikes you as absurd after you wake up. Like this AI driven process really is emulating how our subconscious brains do pattern recognition.
posted by Reyturner at 5:56 PM on September 2, 2021 [1 favorite]

My high point was this thought process:

"Hmmm.... Making out on a beach or something? Wait, is that a tiny Italian flag? Call Me By Your Name!"
posted by kaibutsu at 6:09 PM on September 2, 2021 [1 favorite]

The ones that were obvious to me were obvious at absolute first glance, like looking at the movie from a passing train.

I recognized Miami Vice instantly, but then dismissed the idea because I had forgotten there was a movie version.

Nthing the call for the descriptions. And, I guess, what it was trained on. I'm fascinated that, e.g., Holy Grail included a rabbit.
posted by ChurchHatesTucker at 7:40 PM on September 2, 2021

The four I got: Fear and Loathing in Las Vegas, Mad Max: Fury Road, Point Break and Space Jam. And I still have no clear idea why I got those.
posted by gelfin at 7:44 PM on September 2, 2021 [1 favorite]

The first one I got was Point Break, because that big head in the middle looked so much like Gary Busey. Maybe it's supposed to be Swayze, but it's quintessentially Buseyesque.
posted by Halloween Jack at 7:55 PM on September 2, 2021 [1 favorite]

I got 20 - by the end they felt easy. I guess the examples were training my neural network?
That probably doesn't bode well for my job security, especially now people have figured out all the Copilot swear words.
posted by wilberforce at 7:59 PM on September 2, 2021 [1 favorite]

My partner got 13 right and let me tell you, going through it with them was a super fun Rorschach test kind of game. They also got the "concept" of a lot of ones where they'd never heard of the movie ("That's Florida" = Miami Vice; "That's California" = Fear and Loathing in Las Vegas; "That's a cheearleading movie" = Bring It On, "THAT'S DEFINITELY MARCHING BAND" = Drumline, etc.).

My SIL who's in film school did worse, but I think that worked against her because the reasons she remembers movies are rarely plot-related. She also kept guessing Inception, which didn't help.
posted by brook horse at 8:01 PM on September 2, 2021

The ones I got were instant but I thought a lot of them were supposed to be Blade Runner. (Also Titanic for Mama Mia). I really like how apt the Die Hard one is.
posted by janell at 10:11 PM on September 2, 2021 [1 favorite]

I was fixing to post that I got zero of them but then the Mad Max: Fury Road one came up. I think in total I got five. I agree that Office Space was the most obvious of the lot. Space Jam a close second.
posted by potrzebie at 10:42 PM on September 2, 2021

I find these disturbing in a way I can't quite articulate. The more I stare at them, the more cut off I feel from my sense of self. Maybe that's just me?
posted by misterbee at 11:18 PM on September 2, 2021 [5 favorites]

I don't know why it's only just occurring to me now that, like a lot of AI-generated things, there is probably one final "is this a good example of a thing that is the sort of thing I told you to make?" classifier that they often go through, and that is the human deciding which examples are worth showing, and which aren't. I'd be curious to see what the less guessable results that were nixed are!
posted by Jon Mitchell at 11:29 PM on September 2, 2021 [2 favorites]

I didn't get a single one of these, but looking at the titles, I really like the way the one for Fury Road looks.
posted by entity447b at 12:38 AM on September 3, 2021 [2 favorites]

I find these disturbing in a way I can't quite articulate. The more I stare at them, the more cut off I feel from my sense of self. Maybe that's just me?

It is NOT.

These all make me want to puke buckets, but I got Fury Road and The Hurt Locker, neither of which movies I've seen, before I had to click away.

(Why do so many of them look Pepsi-branded? And why so heavy on the fuchsia? Gurk!)
posted by Don Pepino at 5:20 AM on September 3, 2021 [2 favorites]

The first one, with the steam punk train engine? That’s The Great Train Robbery.

Oh it’s not? So the second time, I got quite a few. Like that’s def Princess Leia in the lower right corner. I never even noticed the pink dresses. Maybe it depends on which part of the image you notice first? And maybe it helps not to have seen the movie.
posted by TWinbrook8 at 6:10 AM on September 3, 2021

Huh, I got eleven. Twelve if my guess of "The Empire Strikes Back" counts for "Star Wars."

I'm supremely fascinated with the area between signal and noise, archetype and mundanity, and these pictures really tickle that sweet spot for me. Supa-nifty!
posted by DrAwkward at 12:05 AM on September 4, 2021

(For generated texts we see online "trained on The Office scripts!" or whatnot, the training done to produce the various things is more fine-tuning the original model than direct training.)


Well, most of the "I made a computer watch/I trained a computer on the scrips of..." texts we see online are fabrications, written entirely by hand and made to sound like a computer because for whatever reason it's funnier if we think a computer did it.

Anyway these are some of the best impressionist works I've ever seen, I'd happily put the Wizard of Oz or Fury Road ones on my wall, and I think it's very funny that a computer did it.
posted by Mr.Encyclopedia at 8:15 PM on September 4, 2021