This project explores the synergy between AI-driven video generation and 3D Gaussian Splatting. By generating a 10-second "time-freeze" helical orbital video, I created a synthetic photogrammetry dataset that allows for a fully navigable 3D scene of an industrial eVTOL (Aircar) and its environment.
Technical Workflow:
- Source Generation: Generated a 10s "Rigid Diorama" video using Seedance 2.0. The prompt enforced a helical spiral orbit with zero motion blur and infinite depth of field to ensure maximum feature tracking for the splatting process.
- Dataset Preparation: Extracted frames at 15 FPS via FFmpeg using high-quality quantization (-qscale:v 1) to preserve 8K-equivalent textures.
- 3D Reconstruction: Trained in LichtFeld Studio using 2M splats over 30,000 iterations.
- Assets: Features a matte carbon fiber Aircar, a high-detail human "Digital Twin".