AniGS from Meta breathes life into static 3D scenes with AI motion
Adds subtle, natural motion like swaying grass to large 3D reconstructions.
Deep Dive
AniGS is a method for adding subtle, distributed dynamics like vegetation motion to 3D Gaussian Splatting reconstructions, preserving rigid structures. It uses a time-conditioned deformation field and a pretrained video diffusion model with an iterative render-and-refine strategy. Tested on five real-world, large-scale outdoor scenes, AniGS produces natural ambient motion and high-quality novel view videos.
Key Points
- AniGS adds realistic ambient motion (e.g., swaying grass, rippling water) to large-scale 3D Gaussian Splatting reconstructions while keeping buildings and roads static.
- Uses a time-conditioned deformation field trained via an iterative render-and-refine pipeline that progressively expands viewpoint coverage across the scene.
- Leverages a pretrained video diffusion model and a composited video-to-video refinement scheme to prevent artifacts in unintended areas.
Why It Matters
Turns static 3D scans into immersive, dynamic environments for AR/VR, gaming, and digital twins.