HBO
House of THE Dragon
We worked with HBO for their House of the Dragon Season 2, marketing campaign. The campaign involved creating AI-generated posters that allowed fans to immerse themselves in the world of Westeros. Fans could scan their faces, choose their allegiance, and see themselves transformed into characters from the series. This interactive effort achieved remarkable engagement, with 8.3 million impressions and 1.8 million unique images generated, sparking excitement for the new season.
As Senior Director of Technology and AI, I had the role of designing the generative image pipeline as well as the face swapping pipeline. During production, I generated and composited half of the final assets, around 250 in the span of three weeks.
USER EXPERIENCE
The user experience was designed to the minimize the number of steps. User’s chose their allegiance, scanned their face, picked their skin tone, and watched themselves be transported to Westeros.
SDXL Training
The project presented several major challenges: building a scalable, fast pipeline to keep hosting costs manageable, producing high-quality assets that aligned with HBO’s standards, and ensuring the experience was inclusive. To meet these demands, I launched an intensive R&D sprint, exploring different Stable Diffusion pipelines. Since the experience required multiple character classes, each with its own unique look, I decided to use LoRA for finetuning. With access to HBO’s production photos, we could create large datasets, and LoRA offered added flexibility for blending character styles later on. In total, I trained 10 distinct LoRAs across various costume sets, resulting in a rich collection of assets that met HBO's technical and creative needs.
DATASET
GENERATIONS
Inclusivity
Because this was a face swap application, we needed to make sure that we could represent as many skin tones as possible. By following Google’s Skin Tone Research, I was able to regenerate every character class with a variety of skin tones.
Asset Creation
In order to drive hosting costs down, the only real time AI functionality was the face swapping. All the backgrounds where generated ahead of time. Each character class had a male and female version, 2 separate poses, and 10 skin tones that needed to be generated. Once those were all generated, they needed to composited with scenery created by HBO’s in house design team.
GENERATION
COMPOSITE
GRADING