Research map / 3D vision / 3D generation
S-INF: Towards Realistic Indoor Scene Synthesis via Scene Implicit Neural Field
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 Scene Implicit Neural Field (S-INF), which disentangles scene layout and detailed object relationships in a latent implicit field, trained with global distillation and differentiable rendering to synthesize realistic, style-consistent 3D indoor scenes retrieved from 3D-FUTURE.
Problem
Learning-based 3D indoor scene synthesis often produces unrealistic layouts (overlaps, misalignments) and style-inconsistent objects because explicit box-based scene representations overlook detailed object information and multimodal relationships.
Approach
Model the latent scene space with an implicit neural field that carries detailed object relationships, plus an explicit spherical layout; distill layout knowledge from a sparse-CNN scene encoder and add differentiable-rendering (normal and mask) losses so the field captures style-consistent object details before CAD retrieval.
Key results
- On 3D-FRONT bedroom with retrieval, S-INF reports FID 79.52 vs. 84.40 for DiffuScene (best baseline in that column).
- Living room FID 78.98 and dining room FID 79.95 with retrieval, both best among compared methods.
- Ablation: disentangled layout plus INF (l_b + f_b) gives bedroom FID 131.47 vs. 156.42 with layout only and 184.86 with INF only.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use S-INF instead of the strongest listed no-retrieval baseline | 3D-FRONT Bedroom · Bedroom, no retrieval | FID: 156.42 -> 131.47 | Helped |
| Use S-INF instead of the strongest listed no-retrieval baseline | 3D-FRONT Living Room · Living Room, no retrieval | FID: 156.48 -> 155.47 | Helped |
| Use S-INF instead of the strongest listed no-retrieval baseline | 3D-FRONT Dining Room · Dining Room, no retrieval | FID: 162.82 -> 157.93 | Helped |
| Use S-INF instead of the strongest listed retrieval baseline | 3D-FRONT Bedroom · Bedroom, retrieval | FID: 84.40 -> 79.52 | Helped |
| Use S-INF instead of the strongest listed retrieval baseline | 3D-FRONT Living Room · Living Room, retrieval | FID: 80.15 -> 78.98 | Helped |
| Use S-INF instead of the strongest listed retrieval baseline | 3D-FRONT Dining Room · Dining Room, retrieval | FID: 84.46 -> 79.95 | Helped |
| Add INF detailed-object relationships to the layout representation (ablation) | S-INF · 3D-FRONT Bedroom | FID: 156.42 -> 131.47 | Helped |
| Add layout relationships to the INF representation (ablation) | S-INF · 3D-FRONT Bedroom | FID: 184.86 -> 131.47 | Helped |
| Add INF detailed-object relationships to the layout representation (ablation) | S-INF · 3D-FRONT Living Room | FID: 156.48 -> 155.47 | Helped |
| Add layout relationships to the INF representation (ablation) | S-INF · 3D-FRONT Living Room | FID: 176.85 -> 155.47 | Helped |
Limitations
- Results rely on CAD retrieval via Chamfer distance, so final object appearance is bounded by the 3D-FUTURE catalog.
- Baselines were modified (3D IoU layout loss, ground-truth mesh supervision) for a fair comparison, which may affect comparability.
Experimental setup
Datasets: 3D-FRONT 3D-FUTURE
Benchmarks: 3D-FRONT
Metrics: FID KL divergence scene classification accuracy diversity
Baselines: Sync2Gen ATISS ScenePrior DiffuScene EchoScene InstructScene
Builds on: DeepSDF NeRF NKField ScenePrior DiffuScene
Closest papers by meaning
- Hierarchically-Structured Open-Vocabulary Indoor Scene Synthesis with Pre-trained Large Language Model (AAAI 2025 · similarity 0.90)
- InstructScene: Instruction-Driven 3D Indoor Scene Synthesis with Semantic Graph Prior (ICLR 2024 · similarity 0.89)
- Direct Numerical Layout Generation for 3D Indoor Scene Synthesis via Spatial Reasoning (NeurIPS 2025 · similarity 0.88)
- SceneCraft: Layout-Guided 3D Scene Generation (NeurIPS 2024 · similarity 0.88)
- InternScenes: A Large-scale Simulatable Indoor Scene Dataset with Realistic Layouts (NeurIPS 2025 · similarity 0.87)
- DeBaRA: Denoising-Based 3D Room Arrangement Generation (NeurIPS 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.