Research map / 3D vision / Face and shape
PrEditor3D: Fast and Precise 3D Shape 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.
Training-free, user-guided 3D shape editing that synchronizes multi-view 2D editing (MVDream + DDPM inversion + Prompt-to-Prompt), segments edited regions via Grounding DINO/SAM 2 lifted to 3D with GTR, and merges voxel features with averaged blending.
Problem
Precise, fast text-and-mask-guided 3D shape editing that changes only intended regions while preserving the rest of the shape; 2D masks are ambiguous after 3D-to-2D projection and optimization-based editing is slow.
Approach
Edit 4 orthogonal multi-view renderings with a multi-view diffusion model, detect intended edit regions in 2D with grounding and segmentation, lift the color-coded segmentation to 3D through a feed-forward reconstructor, and blend original and edited voxel features in 3D so unedited regions stay identical.
Key results
- PrEditor3D reports CLIPdir 1.782 vs. 1.622 for Vox-E and 1.416 for Tailor3D, with the best scores across all directional CLIP variants.
- Runtime is about 74 seconds total (24 s editing, 50 s merging) on an RTX 3090, versus 75 minutes for Vox-E and 6 minutes for MVEdit.
- In a user study with 3600 responses, users preferred PrEditor3D over Vox-E, MVEdit and Tailor3D in prompt alignment, visual quality and shape preservation (e.g., 78%, 88%, 89% vs. Vox-E).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PrEditor3D instead of Vox-E | 3D shape editing · 3D editing evaluation | CLIPdir: 1.622 -> 1.782 | Helped |
| Use PrEditor3D instead of Vox-E | 3D shape editing · 3D editing evaluation | CLIPdir-cos: 6.178 -> 7.679 | Helped |
| Use PrEditor3D instead of MVEdit | 3D shape editing · 3D editing evaluation | CLIPdiff-edit: 10.014 -> 8.812 | Helped |
| Use PrEditor3D instead of Vox-E | 3D shape editing · 3D editing evaluation | CLIPdiff-noedit: 3.432 -> 2.636 | Helped |
| Add 3D merging to the multi-view editing pipeline (ablation) | PrEditor3D · Edited-to-input shape consistency | Chamfer Distance: 2.95 -> 2.28 | Helped |
| Add average blending at edited-region boundaries instead of copy-paste merging (ablation) | PrEditor3D · Edited-to-input shape consistency | Chamfer Distance: 2.29 -> 2.28 | Helped |
Surprising findings
- GPTEval3D does not account for input preservation, so a method that changes shapes globally can still score well on it; the authors therefore added input-aware CLIP metrics.
Limitations
- Limited to the 256x256 resolution of the MVDream multi-view diffusion model.
- Assumes objects can be captured from four inward-facing views, so large scenes such as indoor rooms are not handled.
- Results depend on rough user masks and on the grounding and segmentation models identifying the intended region.
Experimental setup
Datasets: GSO Objaverse
Benchmarks: GPTEval3D directional CLIP metrics (CLIPdir variants, CLIPdiff) user study
Metrics: GPTEval3D preference rate CLIPdir CLIPdir-cos CLIPdir-avg CLIPdir-avg-cos CLIPdiff-edit CLIPdiff-noedit user preference chamfer distance runtime
Baselines: Tailor3D MVEdit Vox-E
Builds on: MVDream Prompt-to-Prompt DDPM inversion GTR Grounding DINO SAM 2
Closest papers by meaning
- Instant3dit: Multiview Inpainting for Fast Editing of 3D Objects (CVPR 2025 · similarity 0.89)
- SHAP-EDITOR: Instruction-Guided Latent 3D Editing in Seconds (CVPR 2024 · similarity 0.87)
- FocalDreamer: Text-Driven 3D Editing via Focal-Fusion Assembly (AAAI 2024 · similarity 0.87)
- Thinking in Boxes: 3D Editing in Real Images Made Easy (NeurIPS 2026 · similarity 0.87)
- Image Sculpting: Precise Object Editing with 3D Geometry Control (CVPR 2024 · similarity 0.87)
- EditSplat: Multi-View Fusion and Attention-Guided Optimization for View-Consistent 3D Scene Editing with 3D Gaussian Splatting (CVPR 2025 · similarity 0.86)
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.