What If: 1975 Studebaker Lark Redux – Making Sausage in a New Way

As promised at the end of my last post about imaginary AI-generated 1975 Studebaker Larks, I’m working on working on what mid-70’s Avantis might have looked like if Studebaker had stayed in business. I’m not thrilled with the results.

While the AI is doing the heavy lifting of actually creating the pictures, what I end up showing you depends on what prompts I give the AI and which images I choose to upscale (render with more detail). And all of that is very much dictated by my own biases, preconceived notions, and expectations.

In any case, I was churning out lots of credible sporty-looking cars but they lacked a clear lineage to the original Avanti. Also, every car had thin 1960’s blade bumpers instead of 5-mph girders or an Endura nose cap.

One way to get around this is to provide a reference image. Thus far I’ve avoided doing that for this project because a) the AI can be too literal and just creates variations of the reference image and b) part of the reason for this project was to see what the AI could do.

However, through working on another project I learned that I could use multiple reference images to achieve the desired result. For example, no matter how hard I tried I couldn’t get it to depict a whistle laying on a blue pillow. However, when I fed it separate images of a whistle and a pillow, I got exactly what I needed, and it didn’t look like it was a copy of either image. Success!

So, in a fit of desperation I tried this technique on the Avanti. In addition to my normal Avanti prompt, I fed it three reference images:

  • An early Avanti
  • A 1975 Chevy Laguna (for the Endura nose)
  • A Porche 928 (for un-chromed nose with visible round headlights)

It came out much more period-correct than my previous attempts. The nose is unique, though I think it would work better in all red. The boy may have gotten too much GM-flavor from the Laguna. In any case, there’s a tease of my next What If post.

The feedback on the AI Larks was varied, ranging from “yeah, I can see that” to “no way.” To each their own. But I thought, why not try the Lark with reference images and see if I can get a better (if maybe less honest) result.

As a baseline, I asked the AI for a 1975 Studebaker Lark without any reference images. It gave me four possibilities.

Of those, I chose to upscale this one. (Note that the upscaling process isn’t just re-rendering the image at a higher resolution. It further refines the image and adds or changes details. Sometimes upscaling adds or removes headlights).

Now let’s try it with reference images. For my first image, let’s use a 1966 Studebaker Cruiser from a Tom Halter Auction Classic post.

To get the correct bumpers and 1970’s aesthetic, I chose a 1975 Chevy Nova from Jason Shafer’s Nova Custom post.

As usual, I got four choices. I discarded the first as looking too much like a Nova from the sides and the third and fourth for having weird shadows covering parts of the body.

So Number Two got upscaled. The usual weird details aside (what are those things on the hood?), how do you think this compares to the Larks from my previous post? Is using reference images gaming the system for these posts?