Research map / 3D vision / 3D generation
GeoEdit: Geometry-Aware Object Editing via Dual-Branch Denoising
Yi He, Jiangming Wang, Xinyu Wang, Mark Fong, Songchun Zhang, Yuxuan Xue, Hai-Tao Zheng, Yue Ma
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add variance-homogeneous injection (vs uncalibrated constraints) (ablation) | GeoEdit · GeoEditBench | DINO: 0.314 -> 0.961 | Improved |
| add warm-start initialization (ablation) | GeoEdit · GeoEditBench | PSNR: 21.320 -> 23.499 | Improved |
| use dual-branch window (47,40) instead of single-timestep init at 47 (ablation) | GeoEdit · GeoEditBench | PSNR: 20.724 -> 23.499 | Improved |
| use dual-branch window (47,40) instead of low-noise init at 40 (ablation) | GeoEdit · GeoEditBench | CLIP: 0.906 -> 0.952 | Improved |
| restrict proxy injection to a window instead of all timesteps (ablation) | GeoEdit · GeoEditBench | DINO: 0.860 -> 0.961 | Improved |
| start from warm-start proxy instead of pure noise (ablation) | GeoEdit · GeoEditBench | PSNR: 18.479 -> 23.499 | Improved |
| avoid excessive proxy injection (t=1) (ablation) | GeoEdit · GeoEditBench | DreamSim: 0.084 -> 0.027 | Improved |
| add variance matching to injection (ablation) | GeoEdit · attention leakage at t=46 | ALR: 8.4% -> 6.4% | Improved |
| use full 81-frame video context instead of 1/21/41/61 (ablation) | GeoEdit · visual ablation | structural integrity | Improved |
| GeoEdit vs NanoBanana on identity (CLIP) | GeoEdit · GeoEditBench | CLIP: 0.976 -> 0.952 | Hurt |
| GeoEdit vs strongest baselines on geometry metrics | GeoEdit · GeoEditBench | Object IoU: 51.3% (Flux-Kontext) -> 57.9% | Improved |
| GeoEdit vs 3DitScene on external benchmark | GeoEdit · 3DEdit-Bench | Object IoU: 0.310 -> 0.460 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- 3DitScene — 3DitScene: Editing Any Scene via Language-guided Disentangled Gaussian Splatting (ICLR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Geometric Image Editing via Effects-Sensitive In-Context Inpainting with Diffusion Transformers (ICLR 2026 · similarity 0.90)
- Training-free Geometric Image Editing on Diffusion Models (ICCV 2025 · similarity 0.90)
- DesignEdit: Unify Spatial-Aware Image Editing via Training-free Inpainting with a Multi-Layered Latent Diffusion Framework (AAAI 2025 · similarity 0.88)
- Thinking in Boxes: 3D Editing in Real Images Made Easy (NeurIPS 2026 · similarity 0.88)
- GeoRemover: Removing Objects and Their Causal Visual Artifacts (NeurIPS 2025 · similarity 0.88)
- Move Anything with Layered Scene Diffusion (CVPR 2024 · similarity 0.87)
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.
