Research map / 3D vision / 3D generation
Infinigen Indoors: Photorealistic Indoor Scenes using Procedural Generation
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.
Extends the procedural Infinigen generator to photorealistic indoor scenes, adding 79 indoor asset generators, a constraint-language arrangement solver, and export to real-time simulators.
Problem
Existing synthetic indoor datasets and generators rely on static asset libraries with limited diversity, and the fully procedural Infinigen system only covers natural scenes.
Approach
A fully procedural (no external assets) indoor scene generator in Blender with procedural furniture/appliance/architecture generators, a domain-specific constraint language for arrangement, and a simulated-annealing solver that optimizes floor plans and object layouts hierarchically.
Key results
- Shadow removal with ShadowFormer: zero-shot SRD non-shadow/overall PSNR improved from 22.83 to 24.56 when adding 2k Infinigen Indoors pairs (R to R+S).
- Occlusion boundary estimation: Infinigen Indoors-trained U-Net reached ODS 29.47 and mAP 19.09, versus 26.02 / 15.69 for Hypersim-trained.
- Solver ablation: full system gives about 3x speedup over the non-optimized version; human study found subjects preferred Infinigen Indoors layouts over prior methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add 2k Infinigen Indoors synthetic images to ISTD training | ShadowFormer · ISTD test | PSNR: 31.96 -> 31.72 | Hurt |
| Add 2k Infinigen Indoors synthetic images to ISTD training | ShadowFormer · SRD test | PSNR: 22.83 -> 24.56 | Helped |
| Train U-Net on Infinigen Indoors images rather than Infinigen-Nature or Hypersim | U-Net · Artist-designed synthetic 3D scenes test | ODS: 26.02 -> 29.47 | Helped |
| Remove discrete changes (ablation) | Infinigen Indoors solver · 20 random scenes, 5k solver steps | Avg. Score (s): 80.84 -> 86.20 | Helped |
| Remove evaluation cache (ablation) | Infinigen Indoors solver · 20 random scenes, 5k solver steps | Avg. Score (s): 80.84 -> 80.84 | No clear effect |
| Remove move filter (ablation) | Infinigen Indoors solver · 20 random scenes, 5k solver steps | Avg. Score (s): 80.84 -> 84.75 | Helped |
| Remove placeholder optimization (ablation) | Infinigen Indoors solver · 20 random scenes, 5k solver steps | Avg. Score (s): 80.84 -> 80.46 | Hurt |
| Remove plane hashing (ablation) | Infinigen Indoors solver · 20 random scenes, 5k solver steps | Runtime: 2280.68 -> 3522.86 | Hurt |
| Remove BVH cache (ablation) | Infinigen Indoors solver · 20 random scenes, 5k solver steps | Runtime: 2280.68 -> 4740.48 | Hurt |
| Run the solver for 100 min. (ablation) | Infinigen Indoors solver · 20 random scenes | Avg. Score (s): 80.84 -> 114.72 | Helped |
| Remove solver optimizations (ablation) | Infinigen Indoors solver · 20 random scenes, 5k solver steps | Runtime: 2280.68 -> 6308.92 | Hurt |
| Reduce solving steps from 10K to 1K (ablation) | Infinigen Indoors solver · Qualitative scene arrangement | scene quality | Mixed |
Surprising findings
- Adding synthetic data slightly reduced in-distribution performance on ISTD while improving zero-shot generalization to SRD.
Limitations
- Human study and realism judgments may be influenced by asset and lighting quality.
- Constraint specification requires user Python effort (about 15 lines per room type in initial spec).
Experimental setup
Datasets: ISTD SRD Hypersim Infinigen-Nature OpenSurfaces
Benchmarks: ISTD SRD Artist-designed synthetic scenes for architecture visualization (occlusion boundaries)
Metrics: PSNR RMSE ODS OIS mAP runtime average score #objects
Baselines: ISTD-trained ShadowFormer Infinigen-Nature Hypersim ProcTHOR ISTD SRD
Builds on: Infinigen Blender ProcTHOR Metropolis-Hastings simulated annealing
Closest papers by meaning
- InternScenes: A Large-scale Simulatable Indoor Scene Dataset with Realistic Layouts (NeurIPS 2025 · similarity 0.85)
- Architect: Generating Vivid and Interactive 3D Scenes with Hierarchical 2D Inpainting (NeurIPS 2024 · similarity 0.84)
- DRAWER: Digital Reconstruction and Articulation With Environment Realism (CVPR 2025 · similarity 0.83)
- S-INF: Towards Realistic Indoor Scene Synthesis via Scene Implicit Neural Field (AAAI 2025 · similarity 0.83)
- InstaInpaint: Instant 3D-Scene Inpainting with Masked Large Reconstruction Model (NeurIPS 2025 · similarity 0.83)
- RoomPainter: View-Integrated Diffusion for Consistent Indoor Scene Texturing (CVPR 2025 · 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.