We developed an AI system capable of generating ads from scratch.
While AI has played a role in digital advertising for some time now, it has largely been relegated to the optimization of existing assets. Someone uploads a spreadsheet of headlines and images, and the AI goes to work mixing and matching them to see what combinations perform best.
But what about coming up with ads from scratch? With the recent launch of Stable Diffusion, which allows us to automate high-quality image generation, we suspected this is now possible. So we decided to test it out.
We created an AI system that combines the story telling capabilities of GPT-3 with the image generation capabilities of Stable Diffusion, and a dynamic Figma template with brand fonts and logos.
The result enables you to generate a fully formed ad in seconds based on a single prompt: “Generate an ad for ________.”
How it works
Synthetic Creative Director
First we set up a “Synthetic Creative Director“ using GPT-3 that generates headline and body copy. To train it we fed it copy from a series of high-quality examples. We decided to stick to ads from the same time period and category to keep it focused — in this case early 2000’s automotive.
To fine-tune the model, we transcribed each ad into a format that GPT-3 can understand and take cues from through language. Here’s an example of the input:
You are an advertising creative director. Come up with an ad for Harley Davidson Motorcycles.
This ad shows a photograph of a blue Harley Wide Glide parked on an open highway at dusk. A classic midwestern landscape of rolling hills stretches as far as the eye can see.
The headline reads: "No. I've decided to lead a rather slow and uneventful life."
The body copy reads: “You could eat up a lifetime pondering what to do with your days on earth. Or you could take one look at a machine like the Wide Glide and let gut instinct take it from there."
Synthetic Art Director
The next step is to take an ad idea like the one above and use it to generate an image prompt for Stable Diffusion. In this case, details matter, so we trained a “Synthetic Art Director” by fine-tuning a GPT-3 model with high quality-examples that we generated by hand.
Here’s an example text output from the training sample, based on the same Harley ad:
Cinematic photograph of a blue Harley Davidson parked on a country freeway, Nebraska prairie landscape, dusk, open skies, horizon, wide angle lens, 8k, hyperrealistic, intricate, award winning photograph
Let’s look at some examples we created with this process 🤖
The following ads were created in response to the prompt: You are an advertising creative director. Come up with an ad for the Tesla Model X.
Curious to learn more?