Research map / 3D vision / 3D generation
DiffuScene: Denoising Diffusion Models for Generative Indoor Scene Synthesis
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 DiffuScene, a denoising diffusion model over unordered object sets (location, size, orientation, class, shape code) for indoor 3D scene synthesis, with shape retrieval to get final geometry.
Problem
Generating realistic, diverse, physically plausible 3D indoor scenes; prior GAN, VAE and autoregressive methods suffer from mode collapse, low fidelity, or sequential error accumulation.
Approach
Represent a scene as an unordered set of objects whose attributes (location, size, orientation, semantics, and pretrained shape-code geometry features) are jointly denoised by a 1D-conv UNet with attention, then retrieve CAD models from the denoised shape codes; an IoU penalty discourages intersections.
Key results
- On unconditional bedroom synthesis, DiffuScene reaches FID 17.21 and KID 0.70 versus 18.60 and 1.72 for ATISS.
- On dining rooms, FID 32.60 vs 38.66 for ATISS; on living rooms, FID 36.18 vs 40.83.
- In a user study over 225 text-conditioned scenes from 45 users, 62% preferred DiffuScene over ATISS for realism.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffuScene rather than ATISS | 3D-FRONT bedroom scene synthesis · 3D-FRONT bedroom | FID: 18.60 -> 17.21 | Helped |
| Use DiffuScene rather than ATISS | 3D-FRONT dining room scene synthesis · 3D-FRONT dining room | FID: 38.66 -> 32.60 | Helped |
| Use DiffuScene rather than ATISS | 3D-FRONT living room scene synthesis · 3D-FRONT living room | FID: 40.83 -> 36.18 | Helped |
| Use DiffuScene rather than ATISS | Text-conditioned scene synthesis · 225 scenes rated by 45 users | user preference for realism: 62% | Helped |
| Replace UNet-1D+Attention with a transformer denoiser (ablation) | DiffuScene · 3D-FRONT bedroom | FID: 17.21 -> 29.08 | Hurt |
| Use multiple prediction heads rather than a single head (ablation) | DiffuScene · 3D-FRONT bedroom | FID: 19.78 -> 17.21 | Helped |
| Add IoU loss (ablation) | DiffuScene · 3D-FRONT bedroom | FID: 17.93 -> 17.21 | Helped |
| Add geometry feature diffusion (ablation) | DiffuScene · 3D-FRONT bedroom | Sym: 0.50 -> 0.72 | Helped |
Surprising findings
- Transformer denoiser (from DALLE-2) performs worse than the UNet-1D with attention used by the method.
Limitations
- Shape retrieval may not match the style of the desired scene since retrieved CAD models are constrained to same-semantic classes.
- Only single-room generation; cannot synthesize multi-room large-scale scenes.
- Relies on 3D labeled scenes; textures come from retrieved CAD models rather than being generated.
Experimental setup
Datasets: 3D-FRONT 3D-FUTURE
Benchmarks: 3D-FRONT
Metrics: FID KID SCA CKL Obj count symmetric pairs PIoU user preference
Baselines: DepthGAN Sync2Gen ATISS LEGO
Builds on: DDPM ATISS Sync2Gen LEGO-Net
Closest papers by meaning
- DiffInDScene: Diffusion-based High-Quality 3D Indoor Scene Generation (CVPR 2024 · similarity 0.90)
- DeBaRA: Denoising-Based 3D Room Arrangement Generation (NeurIPS 2024 · similarity 0.90)
- Move Anything with Layered Scene Diffusion (CVPR 2024 · similarity 0.90)
- Denoising Diffusion via Image-Based Rendering (ICLR 2024 · similarity 0.90)
- FreeScene: Mixed Graph Diffusion for 3D Scene Synthesis from Free Prompts (CVPR 2025 · similarity 0.90)
- LT3SD: Latent Trees for 3D Scene Diffusion (CVPR 2025 · similarity 0.89)
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.