Step 2

Processing Pipeline

All steps run locally using open-source libraries and models. No cloud, no external APIs.

BiRefNet

Background Removal

Each frame is segmented — the subject is isolated from the background using a deep learning model.

COLMAP

Camera Pose Estimation

Structure from Motion reconstructs 3D camera positions and a sparse point cloud from the image set.

Spark

3D Gaussian Splatting

The scene is trained as a set of Gaussians, then exported as a real-time renderable radiance field.