Reproduce “DGE: Direct Gaussian 3D Editing by Consistent Multi-view Editing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Minghao Chen, Iro Laina, Andrea Vedaldi
Researchers at the University of Oxford's Visual Geometry Group developed the Direct Gaussian Editor (DGE), a method for text-guided 3D object and scene editing that utilizes a training-free multi-view consistent 2D editing mechanism with epipolar constraints. This approach enables direct and efficient updates to 3D Gaussian Splatting models, achieving high-fidelity edits in approximately 4 minutes per edit, a substantial speedup over existing techniques.
Key insights
- Preemptively generating multi-view consistent 2D edits for a 3D scene dramatically accelerates the 3D model update process and improves the fidelity of the final 3D output.
- Explicit 3D representations like Gaussian Splatting are inherently well-suited for efficient and controllable localized editing, allowing precise modifications to specific scene components.
- Integrating 3D geometric information, such as epipolar constraints, into the 2D editing pipeline can effectively resolve ambiguities and enhance consistency across multiple views.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.