A few years ago I wrote a series of posts featuring imaginary cars generated by the Midjourney AI. “What if” Packard and Studebaker had survived as US automakers into 1970’s and beyond? Midjourney was happy to render some images based on my prompt, which often looked more like application code than what we expect of AI interaction today. But recent experiments with newer tools allowed me to create a series of AI-generated commercials for cars that never existed. What would a Packard ad be like in the 1970s? If you’re curious, read on.
I haven’t posted anything like the earlier imaginary cars posts because, frankly, Midjourney hasn’t improved much since then (at least, in terms of rendering imaginary cars).
Instead, my current AI fiddling has been focused on the Sora video generator. Given a simple prompt, it will generate a ten to fifteen second video, complete with dialogue, backing music, special effects, and so on. It’s ability to extrapolate and build on what it’s given is sometimes remarkable. Mainly, I’ve focused on jokey videos, such as cats competing the the Winter Olympics, Saturday morning cartoons that never existed, and bad puns.
Then I discovered that Sora could take an image as input. I started uploading those old Midjourney-generated cars (as well as a few newer ones) and asked Sora to create commercials. I only provided a a few guidelines, such as the year, make, model, and type of car. In some instances where Sora was off base I remixed the video with additional guidance on setting, voiceover actor (male, female, educated, British) or whether the car should be in motion. Sometimes it just decided to pan across the still image. Odd.
Often the Sora choices were hilarious and I let them be, like the “tough guy” voiceover in the first Hawk video. How did it come up with that?
When it came to the car names and year, I paid no attention whether I had originally intended to be a Studebaker or a Packard, as I was more intent on seeing what Sora could do rather than be consistent with my earlier posts. I had no notion of writing a post at the time. Eventually, I just started making up names like “Ramses Osiris” and “Fromage Aerodyne.”
Finally, I stitched all of the 15-second videos together into the one you see above. I was never able to get a complete commercial because of Sora’s limit of fifteen seconds, so I treated it as if I were flipping through TV channels and every channel is broadcasting a car commercial.
Meanwhile, over at ChatGPT, their image generation (via DAL-E) has improved greatly, where it is at least on par with Midjourney in terms of realism. The difference is that in ChatGPT a) I can use natural language and b) the AI retains information about what it just did, so I can ask for improvements (except for random remixes of the image just generated, each Midjourney image request is a thing unto itself.
For the image above, I asked ChatGTP, “Imagine if Studebaker had stayed in the automobile business. Then imagine what they might have introduced in 1973 as a personal luxury car named the Hawk. It would be reminiscent of and include styling cues reminiscent of the Hawk of the 1960s, but take into account the market realities of 1973, such as the popularity of personal luxury cars like the Chevrolet Monte Carlo and Oldsmobile Cutlass. Also take into account the US bumper and rollover regulations of that era. Then generate a 16:9 photo of this car from a front 3/4 angle.”
The first image it generated was very much like the one above, except it had quad rectangular headlights. So I asked, “Were rectangular headlights US-legal in 1973?”
ChatGPT gladly gave me a lengthy explanation of US sealed-beam headlight regulations of the era, noted the mistake it had made, and offer to recreate the image with proper headlights.
My request for a 1970 Studebaker Lark was much more problematic, with ChatGPT offering cars with front ends long and wide enough to accommodate a V-8 engine, mid-60’s styling, and generally awkward proportions. My mistake, I think, was continuing the session from where the Hawk was generated, and it hung on to some of those proportions. ChatGPT can hang on to a bad notion like a dog with a bone, and sometimes it’s best to start over.
I started a new ChatGPT session with the request, “Imagine if Studebaker had stayed in the automobile business. Then imagine what they might have introduced in 1970 as an economy car named the Lark. It would be reminiscent of and include styling cues reminiscent of the Lark of the 1960s, but take into account the market realities of 1973, such as the popularity of Japanese imports and new domestic economy cars such as the Ford Pinto, Chevrolet Vega, and AMC Gremlin. Imagine what a Studebaker direct competitor to those cars would be. Then generate a 16:9 photo of this car from a front 3/4 angle.”
Although not exactly a sharp looker, I think this is much closer to what I expected than the previous version.
OK, so I know that I’m probably not supposed to pay attention to the details when it comes to AI generation…but once again I have to ask how come it’s the most simple stuff that it screws up. The numerically reversed, non-linear, out of sequence speedometers that Sora produced are just hilarious.
Thanks for creating these Fuzzyman. It’s good to be reminded that AI still isn’t quite ready to take jobs away from actual car designers. 🙂
There are a lot of weird things. Sometimes it switches cars from left to right hand drive in the middle of the commercial, drives on the wrong side of the road or in a way that seems drunken.
That said, I was kind of impressed with how well it extrapolated what the rear ends of the cars would look like based on the front 3/4 view.
I think the mistakes Sora makes are often the result of it trying to jam so much into such a short video. One of the problems I encounter in videos where people are talking to each other is swapping of voices or dialogue weirdly out of order.
I think that Sora is in video creation kind of where Midjourney was a few years ago in image generation where you had people with too many fingers and such. I’m hoping the next versions of Sora will have similar improvements.
Generally, I found that the extrapolations of the ’50’s models bordered on the ludicrous, but the ’60’s/’70’s ones had some interesting possibilities. That stainless steel, hand inscribed one though, was a rolling circus! Very interesting presentation, Thanks for your efforts!! 🙂
I think some of the 1950’s ones came from my efforts to get Midjourney to generate “the ugliest car in the world,” which would explain a lot.
You got some interesting results; maybe the figuring the stainless steel surfaces like that might hide the usual fingerprint marks!
I had a play with perChance shortly after your earlier posts. I asked for “2001 Riley Kestrel sports saloon deep metallic red on sweeping moorland road’. The LH (part) image was the result. I decided it didn’t look very 2001, had a grille more like a Rootes Singer, far too many, rather randomly placed extra lights on the front and a strange tangle of windscreen wipers. Also the far wing fell away further than the nearer one. I rejigged most of those elements and have given it a title as a 1958 Sunbeam Sports saloon.
I kinda like the 1973 “Hawk”!
One minor bit of data that you might want to specify for future AI play – the 14 inch rims that were prevalent at time!
Is that Packard V8 engine mounted transversely?? Good gawd, what a screwup!
I rather like the clean lines of those late model Studebaker Larks .
-Nate
As a lover of the surreal, I got a real kick out of these videos. As others have noted there’s a slew of absurdities, two of which I took screenshots of, the first one being the gas station attendant filling up the Lark at the base of the windshield:
And this steering wheel missing one of the spokes.
Thanks for this, Fuzzyman. It made my morning!
I like the yellow Lark. Reminds me of a Citation notchback, which wasn’t all that common.
The title picture, all I see is design bits from a bustleback Seville, with most of the body designed for minimal air resistance similar to the DIY AeroCivic that’s been floating around online for about 2 decades.
I definitely want a Fromage Aerodyne!
Isn`t the name ‘Horus’ more appropriate for the Ramses?
AI adds a new definition to cartoons.
The Studebaker Allegro is as magical as its British inspiration, but the Lancia Cortina (mk 3) could have changed automotive history.
My favorite of these is the 1976 Hawk Turbo — one of the cleanest designs without looking too plain.
I noticed that front-seat head restraints are missing from the post-1968 cars.
Here’s something I did on AI……I asked for a blended image for a 1973 GTO as a car from 2026. This is what it showed…..I love it…..comments?