Research map / 3D vision / 3D generation
StageDesigner: Artistic Stage Generation for Scenography via Theater Scripts
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 StageDesigner, a training-free LLM plus layout-controlled diffusion pipeline that turns theater scripts into 3D foreground stage layouts and occlusion-aware 2D backgrounds, and releases the StagePro-V1 dataset of 276 stages.
Problem
Automating artistic stage scenography from theater scripts requires spatial coherence, thematic alignment, and audience-sightline awareness, which existing indoor scene synthesis methods do not handle.
Approach
Decomposes scripts into scene and imagery descriptions, places LLM-generated anchor entities with corner coordinates plus a multi-level collision map, retrieves Objaverse assets, and generates a background via layout-controlled diffusion while excluding regions occluded from the audience's view (foreground projection).
Key results
- Out-of-Bound volume 0.0468 m3 for StageDesigner vs 6.46 m3 for adapted LayoutGPT; OIS 0.756 vs 18.2 m3.
- Class diversity 11.7 vs 7.46 and CLIP-sim 30.3 vs 29.1 over LayoutGPT.
- User study (69 general users, 14 experts) favored StageDesigner over LayoutGPT, with about 72% layout coherence and 73% overall preference votes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Foreground Projection module added (occlusion-aware background bounding boxes) (ablation) | StageDesigner background generation · StagePro-V1 (276 stages) | overlap between background elements and foreground projections: 17.7 -> 2.70 (lower is better) | Helped |
| Background image generation removed (ablation) | StageDesigner · StagePro-V1 (276 stages) | CLIP-sim: 30.3 with background vs 32.1 without | Hurt |
| Script Analysis Module added (ablation) | StageDesigner · 10 scripts user study | Layout Coherence, Overall Preference, Background Fitness (user study): preferred on all three criteria | Helped |
| LLM predicts corner coordinates only for anchor entities; non-anchors placed via Multi-level Collision Map | StageDesigner vs LayoutGPT · StagePro-V1 (276 stages) | Out-of-Bound volume and OIS: OOB 6.46 -> 0.0468 m3; OIS 18.2 -> 0.756 m3 | Helped |
| StageDesigner pipeline (vs adapted LayoutGPT) | stage generation · StagePro-V1 (276 stages) | IWG (m3, higher is better): 9.03 vs 14.5 for LayoutGPT | Hurt |
| StageDesigner pipeline (vs adapted LayoutGPT) | stage generation · StagePro-V1 (276 stages) | Class Diversity and CLIP-sim: Class diversity 7.46 -> 11.7; CLIP-sim 29.1 -> 30.3 | Helped |
| StageDesigner pipeline (vs adapted LayoutGPT) | stage generation · User study (69 general users, 14 experts, 10 scripts) | Layout Coherence and Overall Preference preference share: Layout coherence 72% vs 13% in favor of StageDesigner (first-row count 596); overall preference 73% vs 12% | Helped |
Surprising findings
- Without the background, CLIP-sim was higher (32.1 vs 30.3), which the authors attribute to CLIP limitations rather than worse design.
- LayoutGPT had lower IWG (14.5 m3 vs 9.03 m3 for StageDesigner, where higher is reported as better), which the authors explain by oversized entities.
Limitations
- Evaluation relies on adapted baseline with no existing stage generation methods for direct comparison.
- CLIP-sim is limited by 77-token truncation of long scripts and does not capture atmosphere well.
- Stage size fixed at 1000x1000x1000 cm3 and only a single LLM (gpt-4o) was used.
Experimental setup
Datasets: Objaverse subset (Holodeck, 50,092 assets) LAION-2b StagePro-V1
Benchmarks: StagePro-V1
Metrics: Out-of-Bound volume Overlap-inter-Stage volume Intersection-with-Ground truth volume Class Diversity CLIP-sim user preference
Baselines: LayoutGPT
Builds on: LayoutGPT ReCo Holodeck GLIGEN
Closest papers by meaning
- ArtiScene: Language-Driven Artistic 3D Scene Generation Through Image Intermediary (CVPR 2025 · similarity 0.86)
- RoomPainter: View-Integrated Diffusion for Consistent Indoor Scene Texturing (CVPR 2025 · similarity 0.84)
- SceneX: Procedural Controllable Large-Scale Scene Generation (AAAI 2025 · similarity 0.83)
- Sitcom-Crafter: A Plot-Driven Human Motion Generation System in 3D Scenes (ICLR 2025 · similarity 0.83)
- FreeScene: Mixed Graph Diffusion for 3D Scene Synthesis from Free Prompts (CVPR 2025 · similarity 0.83)
- Architect: Generating Vivid and Interactive 3D Scenes with Hierarchical 2D Inpainting (NeurIPS 2024 · similarity 0.83)
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.