Generative AI systems like GPT-3 are excellent at finding patterns. Sometimes the patterns are obvious, and sometimes they aren’t. That’s useful for generating text for a blog or social post, but you can also use GPT-3’s AI capabilities for your own internal analytics.
To demonstrate this, I used GPT-3 to analyze top-selling stock photos from my company Gado Images’ SFTM collection. This collection focuses on food and travel concepts, largely tied to the San Francisco Bay Area.
These are all Editorial photos, which means they’re limited to uses like news stories–they’re not the often-more-lucrative advertising photos you see on many stock sites.
Analyzing Top-Selling Photos With GPT-3
To start, I went to my company’s DAM and exported a list of our top-selling stock photos from the Sftm collection for the last year. I then made a list of the titles of the top-selling photos, in descending order of yearly sales.
Next, I went to the GPT-3 playground.
I entered the prompt:
I wrote a detailed summary of what these terms have in common
Terms:
1. Bay Area Gas Prices
2. Gott'S Hot Dog And Fries
3. Daily Harvest
4. Starbucks
5. Union Pacific Train
6. Dave And Busters
7. Similac Pro Total Comfort
8. Similac Sensitive
9. Charcuterie Plate
Detailed summary:
These terms are similar because
I used a temperature of .7 and a frequency penalty of .94.
GPT-3 produced this output:
These terms are similar because they are all associated with food and drink. Bay Area Gas Prices is a term typically used to describe the high cost of living in the San Francisco area, while Gott’S Hot Dog And Fries is a popular restaurant chain in the area. Daily Harvest is a food delivery service that offers healthy, pre-made meals and snacks, while Starbucks is a coffee chain that is also known for its food offerings. Union Pacific Train is a mode of transportation that delivers goods across the country, and Dave And Busters is a popular arcade and restaurant chain. Similac Pro Total Comfort and Similac Sensitive are both brands of infant formula designed to provide nutrition for babies. Finally, Charcuterie Plate is a type of dish that typically includes cured meats, cheeses, and other savory foods.
That’s a pretty good summary of the different kinds of photos in the collection that have sold thus far. It illustrates that GPT-3 has some level of understanding of the topic of each photo, and why that topic might have led to sales.
Predicting Future Best-Selling Photos
I wanted to expand beyond understanding what had sold thus far, though, and start to generate ideas for similar photo topics that might also sell well.
To do that, I kept the original prompt output, but expanded it by adding.
Future Stock Photo Ideas:
Building on these successes, we can shoot additional stock photos such as:
From that, GPT-3 produced:
-Food delivery services like Blue Apron or Hello Fresh
-Coffee shops like Dunkin’ Donuts or Peet’s
-Restaurants like Cheesecake Factory or P.F. Chang’s
-Mode of transportation like Amtrak or Greyhound
-Recreational activities like bowling or laser tag
-Baby products like Huggies or Pampers
-Gourmet foods like sushi or filet mignon
That’s a pretty good shot list of related topics. More to the point, it’s a very specific list of topics. I could easily hand that to a photographer and have them go out and shoot those photos.
Discussion
I don’t know how those specific photos would fare in the real world. As a pro photographer, though, they could like plausible high-value photos, at least based on the best sellers in the current collection.
Using GPT-3 to guide photo research could be a great way to come up with new ideas for high-value, top-selling photos. I could see feeding previous winners–or top selling photos within a specific category–into GPT-3 on a regular basis in order to create new shot lots.
It would be great to close the loop and actually take these photos, and then see how they perform. If GPT-3 can follow patterns and predict high-value photos, it would be a real asset.
Photo credit: Photo by Samsung Memory on Unsplash
That is excellent. I am investigating key wording and captions currently, and your piece above is a perfect fit for training the AI on my images. Thank you for the lightbulb moment.