This is the official implementation of "Occam's LGS: A simple approach for Language Gaussian Splatting".
Occam's LGS is a simple, training-free approach for Language-guided 3D Gaussian Splatting that achieves state-of-the-art results with a 100x speed improvement. Our method:
- 🎯 Lifts 2D language features to 3D Gaussian Splats without complex modules or training
- 🚀 Provides 100x faster optimization compared to existing methods
- 🧩 Works with any feature dimension without compression
- 🎨 Enables easy scene manipulation and object insertion
- Training-free global optimization approach
- Direct reasoning in language feature space
- Support for arbitrary language feature dimensionality
- Fast processing time (~15s runtime)
- Compatible with SAM+CLIP features
- Includes tools for object insertion and scene manipulation