The Solution
Our developers worked across the creative core of the platform: turning a prompt into a rigged 3D character, putting that character into real footage, making the movement and cameras reusable afterwards, giving creators an editor to direct the result, and keeping the pipeline underneath it all alive under production load. Six pieces of work carried most of that weight.
From a prompt to a rigged character
A creator types a prompt or uploads an image, and gets back a character that is ready to animate. Our team helped connect and refine that pipeline end to end: generating the concept, converting it into a 3D mesh, applying detailed textures, and rigging the result into a GLB usable anywhere on the platform. Five stages that normally need a modeler, a texture artist and a rigger now run as one pass.
- Generative AI Workflows
- Mesh & Texturing
- Automated Rigging
Characters that inherit a real performance
The platform analyzes uploaded footage, identifies the performers in frame, and replaces them with assigned 3D characters while preserving the movement and structure of the original take. We worked across that flow, from detection through assignment to final output, so a filmmaker can turn live action into character-driven content without recreating every movement by hand in an animation and compositing suite.
- Actor Detection
- Motion Retargeting
- Video to 3D
Movement captured once, reused anywhere
Processing a video does not only produce a finished shot, it produces data. The animation and camera-tracking information generated along the way is stored and can be pulled into new 3D scenes later. Our team helped connect those workflows, so production data stays reusable across the platform instead of being locked inside the single output it was made for.
- Motion Data
- Camera Tracking
- Reusable Assets
The same performance on a different character
Because movement and camera behavior are stored apart from the shot they came from, a creator can take a performance captured in one production, apply it to an entirely different character, drop that character into a newly generated environment, and keep the camera behavior of the original take. World-generation technology is built into the same flow, so those environments do not have to be modeled by hand either.
- Animation Reuse
- World Generation
- 3D Environments
A scene editor for directing the result
Automation gets a creator most of the way. The last stretch needs control. Our team contributed to a full 3D scene editor where users place characters and assets, apply reusable animations, import movement captured from processed video, position cameras, and build cinematic camera paths with keyframes. Finished sequences render straight into Canvas, the platform's node-based environment for AI generation and post-production.
- 3D Scene Editor
- Cinematic Cameras
- Node-Based Canvas
Pipelines that survive their own infrastructure
Frame-level AI work at film resolution is heavy, long-running and full of ways to fail. We rebuilt the processing path so model stages run in parallel across a distributed fleet, cutting end-to-end video processing time by 2.5×, then hardened everything around it: upload and processing automation, codec handling, connection recovery and WebSocket reconnection. A dropped network or a failed node costs a retry, never the job.
- Distributed Compute
- Workflow Recovery
- GPU Scheduling