Reproduce “Free-Editor: Zero-shot Text-driven 3D Scene 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
Nazmul Karim, Hasan Iqbal, Umar Khalid, Jing Hua, Chen Chen
Free-Editor introduces a zero-shot, training-free approach to text-driven 3D scene editing by leveraging a novel single-view editing scheme and a generalized Neural Radiance Field. This method addresses multi-view inconsistency and significantly reduces computational costs, achieving a 20x speedup in editing time while maintaining high 3D consistency.
Key insights
- Editing a single source view and propagating its style through an Edit Transformer effectively resolves multi-view inconsistency issues prevalent in direct multi-view editing approaches.
- The 'self-view robust loss' is crucial for maintaining model generalizability and color consistency across varying sets of source views during inference.
- Leveraging a generalized NeRF architecture enables a truly zero-shot, training-free editing process during inference, drastically reducing computational overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.