Research map / 3D vision / Gaussian splatting
GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Presents GaussianEditor, the first 3D editing method built on Gaussian Splatting, using semantic tracing of Gaussians, hierarchical generation-wise anchor losses, and a 3D inpainting pipeline for object removal and addition.
Problem
Text-driven 3D editing with NeRF is slow and offers limited control over specific regions, and static masks fail as scene geometry evolves; Gaussian Splatting is unstable under highly random generative guidance such as SDS.
Approach
Tag each Gaussian with semantic labels unprojected from 2D segmentation masks so the edit target is tracked through training, and organize Gaussians into generations with increasing anchor-loss constraints on older generations (Hierarchical Gaussian Splatting) to stabilize updates under stochastic diffusion guidance.
Key results
- GaussianEditor-iN2N preferred in 72.28% of user study votes versus 15.45% for Instruct-NeRF2NeRF and 12.27% for GaussianEditor-DDS.
- CLIP directional similarity of 0.2071 for GaussianEditor-iN2N versus 0.1600 for Instruct-NeRF2NeRF.
- A single editing session takes about 5-10 minutes, versus more than 30 minutes for Instruct-NeRF2NeRF; object removal takes about 2 minutes and object incorporation about 5 minutes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GaussianEditor-iN2N instead of Instruct-Nerf2Nerf | GaussianEditor · User study | preference: 15.45% -> 72.28% | Helped |
| Use GaussianEditor-iN2N instead of GaussianEditor-DDS | GaussianEditor · User study | preference: 12.27% -> 72.28% | Helped |
| Use GaussianEditor-iN2N instead of Instruct-Nerf2Nerf | GaussianEditor · Editing comparisons | CLIP Directional Similarity: 0.1600 -> 0.2071 | Helped |
| Use GaussianEditor-iN2N instead of GaussianEditor-DDS | GaussianEditor · Editing comparisons | CLIP Directional Similarity: 0.1813 -> 0.2071 | Helped |
| Add Hierarchical Gaussian Splatting (ablation) | GaussianEditor · Grass-on-fire editing | editing control | Helped |
Surprising findings
- GaussianEditor-DDS produced oversaturated colors and less precise edits than the iN2N variant, indicating guidance choice strongly affects results.
Limitations
- Relies on 2D diffusion models for supervision, which struggle with certain complex prompts.
- Ablation and quantitative evaluation are limited (single user study and CLIP metric, one ablation figure).
Experimental setup
Datasets: multi-view image dataset used for reconstruction of the scene
Benchmarks: user study on editing outputs CLIP directional similarity on editing examples
Metrics: user study preference rate CLIP directional similarity
Baselines: Instruct-NeRF2NeRF GaussianEditor-DDS
Builds on: 3D Gaussian Splatting Instruct-NeRF2NeRF Instruct-Pix2Pix Score Distillation Sampling Threestudio
Closest papers by meaning
- GaussianEditor: Editing 3D Gaussians Delicately with Text Instructions (CVPR 2024 · similarity 0.94)
- EditSplat: Multi-View Fusion and Attention-Guided Optimization for View-Consistent 3D Scene Editing with 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- D-MiSo: Editing Dynamic 3D Scenes using Multi-Gaussians Soup (NeurIPS 2024 · similarity 0.91)
- 3DitScene: Editing Any Scene via Language-guided Disentangled Gaussian Splatting (ICLR 2025 · similarity 0.90)
- Mani-GS: Gaussian Splatting Manipulation with Triangular Mesh (CVPR 2025 · similarity 0.90)
- 3DGS-Drag: Dragging Gaussians for Intuitive Point-Based 3D Editing (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.