Research map / 3D vision / Gaussian splatting
3DGS-Drag: Dragging Gaussians for Intuitive Point-Based 3D Editing
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 3DGS-Drag, a point-based drag editing framework for real 3D scenes that combines copy-and-paste 3D Gaussian deformation with LoRA-fine-tuned diffusion image-to-image correction and progressive multi-step editing.
Problem
Intuitive geometry-related drag editing of real 3D scenes is hard: deformation-based methods need strong priors and 2D-editing-based 3D methods only make appearance or minor geometric changes.
Approach
Deform 3D Gaussians directly from handle/target point pairs (copy-and-paste, keeping originals with lowered opacity), render views, correct them with a per-scene LoRA-fine-tuned DreamBooth diffusion model via inverse-free image-to-image editing with annealed strength, and iterate over progressive drag intervals with handle relocation and history-aware fine-tuning.
Key results
- Editing runs in about 15 minutes with 50 views on a single RTX 4090 (about 2 min fine-tuning, 13 min editing), versus one hour for Instruct-NeRF2NeRF.
- Local editing on the sleeve-extension edit achieves SSIM 0.995 and PSNR 43.43 on unedited regions versus 0.901 and 24.44 without local editing.
- User study and GPT-4o evaluation show 3DGS-Drag outperforms Instruct-NeRF2NeRF and PDS on content correctness and rendering quality.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add local editing (ablation) | 3DGS-Drag · Unedited pixels, “extend sleeves” edit | SSIM: 0.901 -> 0.995 | Helped |
| Add local editing (ablation) | 3DGS-Drag · Unedited pixels, “extend sleeves” edit | PSNR: 24.44 -> 43.43 | Helped |
| Add local editing (ablation) | 3DGS-Drag · Unedited pixels, “extend sleeves” edit | LPIPS: 0.158 -> 0.004 | Helped |
| Remove the local editing mask (ablation) | 3DGS-Drag · 3D scene editing | editing quality | Hurt |
| Use a single drag step instead of progressive steps (ablation) | 3DGS-Drag · Aggressive 3D drag edits | editing quality | Hurt |
| Increase drag steps to 20 (ablation) | 3DGS-Drag · 3D scene editing | editing quality | Helped |
| Use iterative dataset editing instead of annealed dataset editing (ablation) | 3DGS-Drag · Geometry-related 3D edits | editing success | Hurt |
| Use diffusion guidance to correct deformed Gaussians (ablation) | 3DGS-Drag · 3D scene editing | editing quality | Helped |
| Apply 3DGS-Drag instead of Instruct-NeRF2NeRF or PDS | 3D scene editing · 3D scene editing | GPT evaluation and user preference | Helped |
Surprising findings
- Iterative per-frame dataset editing failed to change geometry because inconsistent constraints from unedited views degraded the result.
- A single drag step gave insufficient guidance to the diffusion model, producing broken edits, while more steps only slightly improved quality.
Limitations
- Relies on diffusion model accuracy, so small targets or large complex scenes give suboptimal results.
- Fails on overly aggressive drags that move objects to poorly visible regions.
- Cannot generate unseen sides of objects, since no Gaussians represent them.
Experimental setup
Datasets: Instruct-NeRF2NeRF dataset PDS dataset Mip-NeRF360 Tanks and Temples
Benchmarks: user study on 3D drag edits (19 then 99 subjects) GPT-4o score on edit correctness and rendering quality local editing SSIM/PSNR/LPIPS on sleeve extension
Metrics: user preference GPT score (5-level) SSIM PSNR LPIPS
Baselines: Instruct-NeRF2NeRF PDS SDE-Drag DragDiffusion Deformation baseline
Builds on: DragGAN DragDiffusion SDE-Drag Gaussian Editor Instruct-NeRF2NeRF
Closest papers by meaning
- FlowDrag: 3D-aware Drag-based Image Editing with Mesh-guided Deformation Vector Flow Fields (ICML 2025 · similarity 0.90)
- Dragin3D: Image Editing by Dragging in 3D Space (CVPR 2025 · similarity 0.90)
- EditSplat: Multi-View Fusion and Attention-Guided Optimization for View-Consistent 3D Scene Editing with 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- 3DitScene: Editing Any Scene via Language-guided Disentangled Gaussian Splatting (ICLR 2025 · similarity 0.89)
- GaussianEditor: Swift and Controllable 3D Editing with Gaussian Splatting (CVPR 2024 · similarity 0.89)
- SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis (CVPR 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.