Research map / 3D vision / 3D generation
EditRoom: LLM-parameterized Graph Diffusion for Composable 3D Room Layout 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 EditRoom, an LLM-planned, graph-diffusion-based framework that performs six types of layout edits (rotate, translate, scale, replace, add, remove) from natural language, and introduces the EditRoom-DB dataset with ~83k editing pairs.
Problem
Language-guided 3D indoor scene editing has mostly been limited to appearance changes or requires manual intervention; comprehensive layout edits from natural language commands, including multi-step ones, are not supported, and there is no large dataset for this task.
Approach
An LLM (GPT-4o) parameterizes free-form commands into atomic editing operations, and two graph-transformer diffusion models (target graph diffusion for categories/features/relations, then target layout diffusion for poses and sizes) conditioned on the source scene graph generate the edited scene in a unified way.
Key results
- On the bedroom test set, EditRoom achieves IOU 0.7342 and S-IOU 0.7236, versus 0.7055 / 0.6942 for SceneEditor-N and 0.6220 / 0.6125 for DiffuScene-N.
- EditRoom outperforms both baselines on all four metrics across bedroom, dining room, and living room, and across all six editing types on single-operation commands.
- The model generalizes qualitatively to multi-operation commands without training on multi-operation data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use EditRoom instead of the strongest baseline, SceneEditor-N | 3D room layout editing · Bedroom, dining room, and living room | IOU: 0.7055 -> 0.7342; 0.5262 -> 0.5360; 0.4618 -> 0.4710 | Helped |
| use EditRoom instead of the strongest baseline, SceneEditor-N | 3D room layout editing · Bedroom, dining room, and living room | LPIPS: 0.1261 -> 0.1090; 0.1557 -> 0.1460; 0.1738 -> 0.1616 | Helped |
| use EditRoom instead of the strongest baseline, SceneEditor-N | single-operation editing · Translate, Rotate, Scale, Replace, Add, and Remove | IOU: 0.5983 -> 0.6148; 0.6655 -> 0.6764; 0.6754 -> 0.6881; 0.6451 -> 0.6549; 0.5863 -> 0.6001; 0.6160 -> 0.6399 | Helped |
| use source-scene information as self-attention context rather than concatenating encoded source features with text for cross-attention (ablation) | EditRoom · Bedroom | IOU: 0.5855 -> 0.7342 | Helped |
| use source-scene information as self-attention context rather than concatenating encoded source features with text for cross-attention (ablation) | EditRoom · Bedroom | LPIPS: 0.1432 -> 0.1090 | Helped |
| replace CLIP-ViT-B-32 with OpenCLIP-ViT-bigG-14 (ablation) | EditRoom · Bedroom | IOU: 0.7342 -> 0.6882 | Hurt |
| replace CLIP-ViT-B-32 with OpenCLIP-ViT-bigG-14 (ablation) | EditRoom · Bedroom | LPIPS: 0.1090 -> 0.1305 | Hurt |
Surprising findings
- Using a larger OpenCLIP-ViT-bigG-14 text encoder reduced performance (bedroom IOU 0.6882 vs 0.7342), attributed to limited training data.
- Conditioning on source scene as self-attention context substantially outperforms concatenating it with text features for cross-attention (IOU 0.7342 vs 0.5855).
Limitations
- Performance depends on the LLM's 3D scene understanding and command comprehension, which can yield erroneous or colliding edits.
- Training data is predominantly collision-free, creating a trade-off between following commands strictly and avoiding collisions; extreme commands may be mapped to in-distribution actions.
- Evaluation of multi-operation commands is only qualitative and on manually designed prompts.
Experimental setup
Datasets: 3D-FRONT 3D-FUTURE EditRoom-DB
Benchmarks: EditRoom-DB test set
Metrics: IOU S-IOU LPIPS CLIP score
Baselines: DiffuScene-N SceneEditor-N
Builds on: DiffuScene OpenShape VQ-VAE CLIP 3D-FRONT
Closest papers by meaning
- BlenderGym: Benchmarking Foundational Model Systems for Graphics Editing (CVPR 2025 · similarity 0.88)
- SceneFactor: Factored Latent 3D Diffusion for Controllable 3D Scene Generation (CVPR 2025 · similarity 0.87)
- ConsistDreamer: 3D-Consistent 2D Diffusion for High-Fidelity Scene Editing (CVPR 2024 · similarity 0.87)
- ProEdit: Simple Progression is All You Need for High-Quality 3D Scene Editing (NeurIPS 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 0.87)
- CTRL-D: Controllable Dynamic 3D Scene Editing with Personalized 2D Diffusion (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.