About twenty years ago I captured this image of a unique Dodge Ares while driving through a small town in New Hampshire. Since then, I’ve used it on Christmas cards, social media posts, and it was even used in a holiday CC once.
A few months ago I began experimenting with the Midjourney artificial intelligence image generator. It’s been fairly amazing, and I’ve used it to make images from the style of comic book covers to realistic photos of people and everything in between.
One of the unique things about Midjourney is that images are created by entering a prompt in a Discord chat channel, to which a bot responds with your images. Since it’s a chat channel you can see what images everyone else is making, and the prompts they use to create them (and in the process learn how to create better pictures, or pictures in a similar style).
Well, as I was working on some images, a picture of a rusted out Mustang in the woods very evocative of my Dodge Ares picture scrolled by. So, I copied the prompts for future reference.
And then I went a little nuts.
First, I tried a Camaro. Every time I asked for a picture it would give me four low-resolution options. From those I could ask for more variations, or upscale one or more to a higher resolution.
The results were often odd. Very odd. But along the way I learned a bit about how Midjourney “thinks” and what it “knows.” And what it doesn’t.
(From here on out the captions represent what I asked for, not necessarily what I got).
First, that Camaro. Here were the two best options:
Nice! It’s definitely a first generation Camaro, if not exactly 100% right. (Spoiler, nothing is 100% right). The next one came out more Chevelle than Camaro:
I decided to try something a little more recent: A Chevy Citation. That’s when I ran face first into Midjourney’s knowledge limitations:
Maybe a mid 1960’s Chevelle or Chevy II, at best. And then there was this:
It seems like it didn’t know what a Citation was, and without further information it interpreted “Chevy” as “Big 50’s Chevy with hints of Cadillac.”
I tried a Chevette. No dice. Though I got a bit of a hint of Opel Kadett, which makes sense in a weird way.
Not even close, but the birch trees growing out of the hood are a nice touch.
Maybe if I specified the year? 1979? That was even worse:
Finally, I gave in and tried a 1980 Opel Kadett:
Not bad. Clearly, I was going to have better luck with European cars.
Not the Volvo I expected, but a fair facsimile nonetheless. I asked for a couple of Volkswagens. Without specifying the model, I got a Bug and a Bus.
Quite the VW logo. Throughout using Midjourney I’ve learned that it doesn’t really understand how to represent words or logos. They seem to like a particular type of shape to be mimicked.
I wasn’t sure if Midjourney’s orientation was more European or International, so I asked for a Toyota Hilux.
It got the truck part right, but I’d be challenged to identify that as a Toyota. How about something newer, like a Tesla?
Well done! How about some classic cars, like Studebakers and Packards?
Nope. Generic 50’s car. It looks more like it came from a secret GM division.
As for Packard, this was the only one with something resembling a horse-collar grille.
The rest were more escapees from the GM styling studios:
Back to Europe. How about an Opel Manta?
That’s not any Manta I’ve seen. Maybe an Olds Cutlass?
It certainly says “Oldsmobile” but in more of a C-body way. Also, take note of all the signal lights in the bumper. It seems like to Mindjourney, headlights and such are shapes found on the front of this object, and it doesn’t “understand” that the two sides of the car are usually symmetrical. Most of the pictures I discarded had three (or more) headlights on one side, two on the other. As I asked for more modern cars with composite headlights this became less of a problem.
Taking a different tack, I generically asked for a minivan. One result reminded me of a VW Bus:
Another was appropriately generic.
I couldn’t get a Mercedes grille out of it. Not at all. So I tried a Lincoln Continental, which gave me an interesting amalgamation of Lincoln styling cues from the 50’s and 60’s. It’s a Lincoln, but not any Lincoln that ever existed.
I was starting to figure out the cars had to be iconic to get a good picture, like this Land Rover:
…or this Jeep. Too many slats. Maybe it’s a knockoff.
Of course, I had to try my own Mustang picture:
A request for a Ford F-150 got me some interesting results. Older:
It’s not exact, but it made a Fiat recognizingle:
The same with a Deloran. About half of the Delorean images has Back to the Future accessories. Also, there’s that headlight problem, again.
It got completely confused by the concept of a PT Cruiser:
But a BMW was no problem… and those headlights.
I wanted to see if Midjourney could depict a 2004 SsangYong Rodius, considered by some to be the ugliest car ever made. Alas, the AI couldn’t bring itself to create such a horror and gave me a generic minivan instead:
A Pontiac Grand Prix wasn’t sufficiently famous, so I ended up with something that looked more like a Cutlass:
But a 1966 GTO? No problem!
The picture of a 1964 Buick Rivera came out more like what would have happened if they adapted Riviera styling cues to a C-body. Interesting, really.
To round things out I asked for a Dodge Ares K-Car and got a 1930’s Dodge instead:
Also, maybe a kind of Valiant?
Finally, I remembered that you can give Midjourney a reference image, so I used added a link to my Dodge Ares image to my previous prompts and specified the year 1983 and to give it a fighting chance. The first attempt looked more like an 1980 Newport:
But the last one came out well. We’ve come full circle. The greenhouse is a bit weird and the Mad Max bumper guard is an interesting choice, but I like it.
With car identification being one of my oldest, most acute, and prideful capabilities (well up to a period when cars had more distinct shapes), any real encountering of vehicles resembling those in this post would make me think I was either going insane, had already gone insane, or was succumbing to a form of dementia.
Not that there’s a big difference between sanity and insanity… is there?
Not so much insane as just dreaming, for me. I dream cars like this a lot; and in the dream I “know” what the car is, but it always looks quite a bit different than what it’s supposed to look like, very much like these. In my dream, I say to myself “I don’t remember VW buses having so many side windows, or headlights that bulge out like that”. Or so on, including cars that don’t look hardly at all like the real thing, as some of these do.
Some/many of my dream cars are even more outlandish, incorporating parts of airplanes and/or having the ability to fly, with me at the controls: “how could I have forgotten that cars can fly?”
Despite the winter-ish scenes in a lot of them, most of these cars should have been done in late October. Definitely a Halloween feel to a lot of them.
Wow, I definitely need to try this.
You could be their beta tester…or is this the “stable” version already?
Having been a beta tester… unplanned… for some “wonderful” implementations hatched by my employer, at least I can say I was being PAID!
Some very nice images created. I’ve found AI seems to know the very more iconic cars but not much else. Not particularly surprising though. The creations are quite interesting however. I quite like its “Mercedes” – it is almost a combination of British and American styling. The “Generic 50’s car” is quite attractive as well.
I’m working on getting the AI to create images of what independents like Studebaker would be selling had they survived to 1975 and beyond. I’ve had some interesting results and it’s easy to give the AI more credit than it deserves for the choices it makes. I’ll probably post something in January, but here’s a preview.
Nice! And there’s those multiple rear windows again. It’s really hung up on that, eh?
It really is! Like the headlights it gets the idea that there should be some of these things but doesn’t understand the logic behind them.
It would probably still be another facelift of the 1953 body…
I’m quite impressed; more so than the ones we’ve seen before here. I really like the way certain small details have been changed, like the bulging headlights on the VW bus and Beetle, and the slightly different front turn signals on the VW bus, and so forth.
But I also like the ones that don’t look anything (or very little) like they’re supposed to. Or in between, like the ’64 Riviera sedan, which looks very much like a Cadillac that never got built. Considering that the Riviera was supposed to be a Cadillac, that’s a rather clever outcome, one which I assume was not quite intentional.
Many of these are very convincing; if you showed them to someone not knowledgeable, I would assume they would think these were “real”.
I find it curious that it has a strong tendency to add one or more side windows. Even the iconic Beetle gets an extra side window; curious.
Thanks; I really enjoyed this trip. As I said in an earlier comment, these are a lot like cars I encounter in my dreams, and I always find myself saying “why did I think these were supposed to look different?” or “why did I forget that this is what these really looked like” or some such thing.
Thanks! Getting what you want out of Midjourney is almost like like writing a short computer program. It’s costs money to use (thankfully, not a lot) but the range of what it can make seems to be limited only by how well I craft my prompts.
I’m certainly showing my age here, although I’m old enough to be a little scared of stuff like this (AI stuff), but young enough that I “should” be embracing it?! Either way, I find myself learning/knowing just enough to get me along with my I-phone and computers, but not venturing off into this space.
Kind of makes me want the old days back when we weren’t being inundated with a cell phone every 30 seconds between emails/text/calls, where we still had to drop a roll of film into the old 35MM camera. I don’t know about the rest of you, but I have a good number of photos on my phone that I have no clue what to do with them. I’m sure I could get them off the “cloud” if my phone was lost, but that seems like a lot more work than just grabbing the photo album.
Some of those are unrecognizable and some would be improvements on the original. Fascinating.
Very interesting images.
The Opel Manta reminded me more of Lancia Beta coupe, the first Minivan looked like a Fiat 900T rear engine van.
Italian design influence?
I like many (? most ?) of these, indeed there’s a lot of GM in that program .
I can see this being used to create vehicles for movies or book illustrations, once you get the prompts figured out it’ll be great .