Reproduce “Geometry-Aware Style Transfer in 3D Gaussian Splatting” 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
Min Hyeok Bang, Jun Hyeong Kim, Seung-Wook Kim, Se-Ho Lee
A framework for 3D Gaussian Splatting introduces geometry-aware style transfer, enabling simultaneous modification of a 3D scene's appearance and geometric structures. The method uses a decoupled optimization scheme and multi-modal feature matching to achieve spatially and geometrically coherent stylization across novel viewpoints, outperforming prior 3DGS approaches in style fidelity and multi-view consistency.
Key insights
- Decoupling the optimization of color and geometry parameters is crucial for stable and coherent 3D style transfer within 3DGS, resolving issues of parameter interference.
- Integrating multi-modal features (RGB, depth, edge) into a contrastive learning objective provides a robust mechanism for capturing and transferring distinctive structural characteristics in geometry-aware stylization.
- A geometry-dominant update schedule within the decoupled optimization cycle is beneficial for achieving expressive and stable 3D style transformations, as indicated by ablation studies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.