1.3 million videos, one for each customer.
Carvana's Joyride campaign gave each customer a video about the day they bought their car: their name, the car, the date and place, and what was happening in the world that day. AI wrote the scripts and generated the images and voice, and a serverless render pipeline on AWS produced up to 300,000 videos an hour, more than 1.3 million in total.
Most of the work came first. From March 2022 we spent close to a year finding what such a video could look like, down four paths: Disco Diffusion series such as Gummy car and Meadow car, GANs with Pytti, car-to-landscape morphs in Houdini rendered in Mantra, and point clouds built from photography, with further tests in Unreal. The finished films build on Carvana's own patented vehicle photography.
Data
Name, car, date and place of each purchase.
Script
AI writes a script with what was happening in the world that day.
Generate
Images and a trained voice are generated for each story.
Render
A serverless pipeline on AWS renders up to 300,000 videos an hour.
Working on something like this? Tell us what you need.
