In a bin in the garage, I recently found a mini-poster for the local radio station in Asheville, North Carolina when I was in college. Somewhere I have a photo of my freshman dorm room with it on the wall.
I’ve always been fascinated by this odd piece of art. It has a bit of a 1980’s synthwave feel (and if you lived in the early 90’s, they weren’t that different). Rather than being a random illustration, it actually shows a small city view, likely meant to be Asheville or nearby Greenville, South Carolina, from a tall mountaintop with radio antennas.

Besides being almost 35 years old, the original poster had ripped, been taped, had wear on the edges, and had been rolled up at one point in a humid area that caused the paper to rib. There was no easy way to restore it to its famous glory.
I did, however, take a photo of it with my phone and make some improvements in Photoshop to fix the wear and tear. Not bad, but still, I wasn’t satisfied. I wondered what AI might make out of it if I asked it to remake it.

My first thought was Adobe Firefly because I knew that you could upload an image as a reference, which is a lot easier than explaining a nameless art style in a text prompt. I think this really gives Firefly an edge. I used this functionality to make the images for the AccuWeather article “AI weather photos: How to tell fact from fiction” that I wrote earlier this year (because it was a gray area whether we could legally reuse someone else’s AI images in our article). There’s even a slider to make it less/more like the original.
Once I had added the original as a Composition and Style reference, I gave it this prompt:
An art poster for a radio station in 1992. The scene shows a city at night, mountains with radio antennas in the foreground. The point of view is from above the mountains which are above the city, with a pink sunset and full moon. The downtown is lit up as a bright glow with skyscrapers with lights around it, the rest of the city shows only occasional lights. The radio station’s name “Power 93” is shown in green, overlaid on the city with a lightning bolt icon below the text.
I generated 7 sets of 4 images, 28 total, and out of those I picked 3 great ones. The hardest thing (as is true with most early AI models) was spelling “Power 93” correctly. It nailed the background idea from the get-go, though it was more detailed than the original poster, and went a bit into a Japanese art at times… I think.

For the last two generations, I realized that the original had yellow in the font as well as green, and I clarified that the font should be overlaid on the scene, so it wouldn’t get caught up in trying to place it over the city. I also realized I should have specified that it was a “brushstroke” font; prior to that it only managed to partially guess that based on the original. The changes were:
“Power 93” is shown in green, overlaid on the city with a lightning bolt icon below the text.
“Power 93” is shown in a brushstroke green and yellow font, overlaid on the scene with a lightning bolt icon below the text.
The very first image it generated was almost perfect, except the lightning bolt was off the bottom of the image and there was a weird shadow artifact above it (both of which could be easily fixed in Photoshop). All in all though, great job. It even added the moon’s face, which was not something I asked nor something in the original.

Another version that I enjoyed was this next one, where it made (without being asked) the lightning bolt kind of 3D and silver, which was a nice touch and probably something Firefly had seen frequently during its training. It also got the “brushstroke font” right in a way that I couldn’t have even explained easily in the prompt. The mountains in the background were a nice addition too, and honestly were missing from the original poster! Both the Asheville and Greenville area would likely have mountains in the background as well as foreground.
I almost wouldn’t change anything in this image except separating the “o” a bit from the “p.” At first I was mad that it mixed seasons on the mountainsides in the foreground, until I realized that the mountainside on the right might have been meant to be overexposed because the light source on the right side of the 3D lightning icon would have shone on it alone.

Another of my favorites was this image. The only problem here was that it tried to connect the “o” to the “w” in cursive, making it look like an a, as in, “Pawer 93.” I really like the way it did the city and foreground too, almost adding another 3D layer between the mountain and the city that serves as a kind of suburb.

At this point I had spent 7 of my 25 monthly Firefly points on this project, so I decided to call it quits and move over to Chat GPT 4. Even though though Chat GPT doesn’t have an option for a “reference” image, I uploaded the original anyway and said “with the same style and vibe from this image.”
At this point I was questioning the amount of time I had spent on this project in general, so I spent less time in Chat GPT, only generating two (ish) images, but I quickly got a great result. In general I believe that Chat GPT is more powerful and more accurate than Adobe Firefly, but the interesting thing with it was that it was less artistic, and it makes sense that Adobe would have an AI that would be more like an artist interpreting your request.

Chat GPT also got hung up on my “1992” specification which I realized I shouldn’t have said in the prompt, especially since the art was not really in a 1992 style, and even if it were, the AI wouldn’t understand that. So, after the first image I omitted that from the prompt. The second image was almost perfect, but (as Chat GPT does) it tried to create a vertical and square image, even though the reference image was horizontal. I knew once I saw the second one that I had a winner, so I asked it to redo it as 16:9.

Although it doesn’t have the same foreground idea as the original, this is pretty much spot-on for what I would call a “modern recreation.” It’s more detailed than the original, but not as “chaotic” as the Firefly examples. Because it’s more simple, there’s less room for artistic error. I think this is the winner. Thanks, Chat GPT!
By the way, the radio station in question is still operating in the Asheville/Greenville area. It’s WTPT 93.3 FM. Wikipedia says their claim to fame in 1991 was that they flipped to contemporary hit radio, so it’s no wonder they were a favorite of my freshman dorm that year. It was then that they adopted the moniker “Power 93” – though they abandoned it in August 1992 for “93.3 WBBO” as they became “dance oriented.” They switched to country, rock and gospel over the next several years, and it wasn’t clear what type of music they play today.