Whitehat · 2024
ACT 3 AI
A pipeline at Whitehat that turned user data into finished character videos without an artist touching Blender. Data became YAML patterns, patterns became scenes, and a farm of office PCs rendered them around the clock.
- Role
- Software engineer
- Year
- 2024
- Origin
- Whitehat
- Stack
- Blender · Python · Temporal
01
What it is
ACT 3 AI generated personalized 3D character videos at scale. Instead of someone assembling every video by hand, the system derived YAML patterns from user data and used them to build Blender scenes automatically: characters, props, cameras, and timing, all described in the pattern.
The characters themselves were animated with full-body motion capture, and their mouths were driven straight from the audio track, so a new script meant a new video, not a new animation pass.
02
What I built
- 01The pattern-to-video pipeline: YAML in, rendered Blender video out, with ffmpeg handling assembly at the end.
- 02Character animation: full-body motion capture retargeted onto 3D characters, with lip-sync driven by the audio itself.
- 03Distributed rendering: a master-worker setup on Temporal that turned on-prem office PCs into compute nodes, with job state and retries handled by the workflow engine instead of hand-rolled scripts.