Research map / 3D vision / Gaussian splatting
ProEdit: Simple Progression is All You Need for High-Quality 3D Scene 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 ProEdit, which decomposes instruction-guided 3D scene editing into difficulty-aware progressive subtasks (via text-embedding interpolation) and edits a 3DGS scene with an adaptive Gaussian creation strategy and iterative dataset updates using Instruct-Pix2Pix.
Problem
Instruction-guided 3D scene editing via 2D diffusion distillation suffers multi-view inconsistency because the diffusion model's large feasible output space (FOS) for an instruction makes the scene representation fail to converge to edited views, producing blurred textures and noisy geometry.
Approach
Control FOS size by decomposing an editing task into subtasks defined by interpolating the instruction text embedding toward the empty prompt, adaptively choosing subtask ratios by an LPIPS-based difficulty metric, and applying them progressively with an adaptive 3DGS that resets opacity and throttles Gaussian creation per subtask.
Key results
- On the ScanNet++ comparison, ProEdit achieves USO 87.96 vs 68.65 for ConsistDreamer and 51.35 for IN2N, with US3D 80.23 and GPT 81.00.
- ProEdit reaches comparable CLIP scores to ConsistDreamer (CTIDS 0.0844 vs 0.0912; CDC 0.3833 vs 0.3912) at about 1/3 of its running time (1-4h vs 12-24h).
- Ablation removing subtask decomposition ('ND') lowers USO from 92.70 to 68.46 and US3D from 90.48 to 61.72 relative to the full method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use ProEdit instead of ConsistDreamer | 3D scene editing · Editing evaluation | USO: 68.65 -> 87.96 | Helped |
| use ProEdit instead of ConsistDreamer | 3D scene editing · Editing evaluation | US3D: 75.23 -> 80.23 | Helped |
| use ProEdit instead of ConsistDreamer | 3D scene editing · Editing evaluation | GPT: 74.40 -> 81.00 | Helped |
| use ProEdit instead of ConsistDreamer | 3D scene editing · Editing evaluation | CTIDS: 0.0912 -> 0.0844 | Hurt |
| use ProEdit instead of ConsistDreamer | 3D scene editing · Editing evaluation | CDC: 0.3912 -> 0.3833 | Hurt |
| use ProEdit instead of ConsistDreamer | 3D scene editing · Editing evaluation | Running time: 12-24h -> 1-4h | Helped |
| add subtask decomposition (ablation) | ProEdit · Editing evaluation | USO: 68.46 -> 92.70 | Helped |
| add subtask decomposition (ablation) | ProEdit · Editing evaluation | US3D: 61.72 -> 90.48 | Helped |
| add subtask decomposition (ablation) | ProEdit · Editing evaluation | USP: 60.73 -> 88.72 | Helped |
| add subtask decomposition (ablation) | ProEdit · Editing evaluation | GPT: 72.87 -> 82.80 | Helped |
| add subtask decomposition (ablation) | ProEdit · Editing evaluation | CTIDS: 0.0671 -> 0.0844 | Helped |
| add subtask decomposition (ablation) | ProEdit · Editing evaluation | CDC: 0.2902 -> 0.3833 | Helped |
Surprising findings
- Removing subtask decomposition causes the largest drop across all user study metrics, indicating decomposition matters more than the add-ons in prior methods.
Limitations
- Constrained by the underlying 2D diffusion model's editing capability and lacks 3D awareness, so multi-face/Janus problems may occur.
- Relies on 3DGS, inheriting its limits such as poor suitability for unbounded outdoor scenes.
- Requires more subtasks and total running time than a single-step edit.
Experimental setup
Datasets: IN2N dataset ScanNet++
Benchmarks: IN2N dataset scenes ScanNet++ scenes
Metrics: USO US3D USP GPT score CTIDS CDC running time
Baselines: Instruct-NeRF2NeRF ConsistDreamer ViCA-NeRF CSD PDS Efficient-NeRF2NeRF DreamEditor
Builds on: Instruct-NeRF2NeRF Instruct-Pix2Pix 3D Gaussian splatting SDEdit Splatfacto
Closest papers by meaning
- EditSplat: Multi-View Fusion and Attention-Guided Optimization for View-Consistent 3D Scene Editing with 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- CTRL-D: Controllable Dynamic 3D Scene Editing with Personalized 2D Diffusion (CVPR 2025 · similarity 0.90)
- DreamCatalyst: Fast and High-Quality 3D Editing via Controlling Editability and Identity Preservation (ICLR 2025 · similarity 0.89)
- ConsistDreamer: 3D-Consistent 2D Diffusion for High-Fidelity Scene Editing (CVPR 2024 · similarity 0.89)
- 3DitScene: Editing Any Scene via Language-guided Disentangled Gaussian Splatting (ICLR 2025 · similarity 0.89)
- Instruct 4D-to-4D: Editing 4D Scenes as Pseudo-3D Scenes Using 2D Diffusion (CVPR 2024 · similarity 0.88)
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.