Research map / 3D vision / Gaussian splatting
EditSplat: Multi-View Fusion and Attention-Guided Optimization for View-Consistent 3D Scene Editing with 3D 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.
Proposes EditSplat, a text-driven 3D Gaussian Splatting scene editor that uses multi-view fusion guidance in InstructPix2Pix's classifier-free guidance and attention-guided pruning/selective optimization of Gaussians.
Problem
Text-driven 3D scene editing with 3DGS suffers from multi-view inconsistency in 2D diffusion edits and inefficient optimization because pre-trained Gaussians retain too much source information.
Approach
Fuse depth-projected multi-view edited images into the diffusion guidance of InstructPix2Pix (as an extra classifier-free guidance term alongside source image and text), and use cross-attention maps lifted onto Gaussians to prune high-attention Gaussians before training and restrict optimization to high-attention ones.
Key results
- EditSplat achieves CLIPdir 0.1431 vs 0.1222 for the best baseline DGE.
- EditSplat achieves CLIPsim 0.2531 vs 0.2394 for DGE.
- In a 100-participant user study on Mechanical Turk, EditSplat is preferred 0.4227 of the time vs 0.2143 for DGE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Multi-view Fusion Guidance (MFG) (ablation) | EditSplat · Text-driven 3D scene editing | Multi-view consistency | Helped |
| Add Attention-Guided Trimming (AGT) (ablation) | EditSplat · Text-driven 3D scene editing | Editing quality | Helped |
| Add Attention-Guided Trimming (AGT) (ablation) | EditSplat · Text-driven 3D scene editing | Semantic localization | Helped |
| Use EditSplat instead of DGE | Text-driven 3D scene editing · 8 scenes, 2 prompts each | CLIP text-image direction similarity: 0.1222 -> 0.1431 | Helped |
| Use EditSplat instead of DGE | Text-driven 3D scene editing · 8 scenes, 2 prompts each | CLIP text-image similarity: 0.2394 -> 0.2531 | Helped |
| Use EditSplat instead of DGE | Text-driven 3D scene editing · 8 scenes, 2 prompts each | User study preference: 0.2143 -> 0.4227 | Helped |
Limitations
- Relies on the quality of the 2D diffusion model (InstructPix2Pix) for complex or nuanced prompts.
- Projection quality depends on 3DGS rendering and depth map quality.
- Evaluation relies on CLIP metrics and a user study on a limited set of 8 scenes with 2 prompts each.
Experimental setup
Datasets: IN2N scenes BlendedMVS Mip-NeRF360
Benchmarks: IN2N scenes BlendedMVS Mip-NeRF360
Metrics: CLIP text-image directional similarity CLIP text-image similarity user study preference
Baselines: GaussianEditor GaussCtrl DGE
Builds on: InstructPix2Pix 3D Gaussian Splatting InstructNeRF2NeRF GaussianEditor DreamEditor
Closest papers by meaning
- SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis (CVPR 2025 · similarity 0.92)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 2025 · similarity 0.91)
- GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting (CVPR 2024 · similarity 0.91)
- 3DitScene: Editing Any Scene via Language-guided Disentangled Gaussian Splatting (ICLR 2025 · similarity 0.90)
- ProEdit: Simple Progression is All You Need for High-Quality 3D Scene Editing (NeurIPS 2024 · similarity 0.90)
- Quantifying and Alleviating Co-Adaptation in Sparse-View 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
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.