Research map / 3D vision / 3D generation
GraphDreamer: Compositional 3D Scene Synthesis from Scene Graphs
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 GraphDreamer, which generates compositional 3D scenes from scene graphs (or text converted to scene graphs by ChatGPT) by optimizing per-object SDF fields with node- and edge-level score distillation sampling.
Problem
Text-to-3D methods generate holistic models from plain text and struggle with multi-object scenes, suffering from attribute confusion (attributes bound to wrong objects) and guidance collapse (parts of the prompt ignored).
Approach
Decompose a scene graph into global, node-wise and edge-wise text prompts; represent each object with an identity-aware hash-grid SDF field and render objects individually, so SDS gradients from each object or edge prompt update only the relevant objects, with penetration and Eikonal constraints for plausible geometry.
Key results
- GraphDreamer achieves mean CLIP score 0.3326 vs 0.3267 (Magic3D 0.3102... see table: Magic3D 0.3267, MVDream 0.3102, w/o graph 0.3019) on 30 multi-object scenes; full model 0.3326 is highest.
- Ablation removing the scene graph lowers mean CLIP score by more than 3 std over full model, indicating guidance collapse.
- Decomposition analysis: CLIP score with own prompt 0.308 vs 0.201 with other prompts, showing objects are separated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add scene-graph guidance with node-wise and edge-wise prompts (ablation) | GraphDreamer · 30 multi-object scenes (number of objects ≥2) | CLIP score (mean): 0.3019 -> 0.3326 | Helped |
Limitations
- Relies on scene graph quality; ChatGPT-generated graphs may be wrong
- Quantitative evaluation limited to 30 scenes with CLIP score only
- Compute and optimization cost per scene (20K steps)
Experimental setup
Datasets: Visual Genome-style scene graphs (prompting format)
Benchmarks: 30 multi-object text prompts (CLIP score evaluation)
Metrics: CLIP score mean std
Baselines: Magic3D MVDream GraphDreamer (w/o graph)
Builds on: DreamFusion Magic3D NeuS Instant NGP MVDream
Closest papers by meaning
- R3CD: Scene Graph to Image Generation with Relation-Aware Compositional Contrastive Control Diffusion (AAAI 2024 · similarity 0.89)
- Disentangled 3D Scene Generation with Layout Learning (ICML 2024 · similarity 0.89)
- Towards Text-guided 3D Scene Composition (CVPR 2024 · similarity 0.89)
- 3D-SceneDreamer: Text-Driven 3D-Consistent Scene Generation (CVPR 2024 · similarity 0.88)
- Scene Graph Disentanglement and Composition for Generalizable Complex Image Generation (NeurIPS 2024 · similarity 0.88)
- Object-Centric Representation Learning for Enhanced 3D Semantic Scene Graph Prediction (NeurIPS 2025 · 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.