Prompt Architecture for Scalable Image Generation
Google Hackaton
In early 2024, I participated in a global internal hackathon organized by Google. The goal was to experiment with their latest generative AI tools, Imagen 3 and Gemini and come up with a concept that could improve or reinvent a creative process using those technologies. Several Critical Mass teams from different regions joined, and I was part of the LATAM team, working alongside four other teammates from different disciplines.
Our idea was to build an automated content system for a fictional pet food brand. The core of the project was generating consistent, high-quality images of pets that could be used across social and display campaigns, not just random pets, but the same pet doing different things, in different scenarios. That consistency was key for the concept to make sense.
My two main focuses were helping create the brand identity and figuring out how we could get Imagen 3 to generate images that felt coherent. That meant designing a prompt system, one that could be tweaked in very specific ways to control things like breed, color, action, angle, and background, while still keeping the pet recognizable across all versions. This was new territory for me, and building that system took a lot of testing and iteration, but it worked. With a well-structured prompt, we could create a bunch of variations of the same dog or cat by just swapping out certain pieces.
We had just one day to execute the full thing. The week before was used for light planning and ideation, but the entire build and presentation happened during the hackathon. In the end, we delivered a working prototype and presented it to the other teams and a panel of judges that included people from Google. It was a super intense day, and the fact that we were able to make it all work with solid branding, a working image generation system, and live creative output was something I was really proud of. I got great feedback from my team and the judges, especially around the prompt structure, which turned out to be a key piece of the whole project.
Today, building something like this is already a lot easier. AI tools have evolved quickly, and many of the things we were solving back then now feel almost trivial. But at that moment, it was all about figuring things out as we went, and pushing the limits of what was possible with the tools available.