Building Rome from a Single Image
10 hours ago
- Interactive reconstruction viewer with orbit, zoom, WASD movement, and distance measurement.
- Meshes simplified to ~1.5M faces for web viewing; paper uses full-resolution meshes.
- Scene generator uses adaptive chunks that vary size based on distance to preserve detail.
- Explicit 2D–3D conditioning lifts image features onto surfaces and labels voxels as free, visible, or hidden.
- Autoregressive generation reuses overlapping latents across neighboring chunks for scene consistency.