Research map / 3D vision / 3D generation
DeBaRA: Denoising-Based 3D Room Arrangement 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.
Proposes DeBaRA, a lightweight EDM-based score model that generates 3D room layouts (oriented boxes given floor plan and categories), plus a Self Score Evaluation procedure to select LLM-proposed category sets.
Problem
Generating realistic, diverse and bounded 3D indoor furniture layouts, where existing autoregressive or off-the-shelf diffusion methods entangle categorical and spatial attributes and often violate room bounds.
Approach
Restrict the diffusion output space to spatial attributes only (position, rotation, dimension of oriented 3D boxes), condition on floor plan and category set, train with EDM parameterization and a semantic-aware Chamfer loss plus conditioning dropout, and use the unconditional density to score LLM-proposed category sets (SSE).
Key results
- On 3D layout generation (living rooms), DeBaRA reports FID 18.89 and OBA 167.8 versus DiffuScene 21.54 / 341.1 and ATISS 25.67 / 857.3.
- On 3D scene synthesis, LLM + SSE improves living-room FID from 21.58 to 20.59 and OBA from 154.3 to 152.0 over random LLM selection.
- Sampling with EDM 2nd-order sampler (25-50 steps) takes about 0.25-0.5 s versus about 5 s for 1000-step DDPM, with same-or-better FID.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeBaRA instead of DiffuScene for bounded living-room layout generation | 3D layout generation · 3D-FRONT living rooms | FID: 21.54 -> 18.89 | Helped |
| Use DeBaRA instead of DiffuScene for bounded living-room layout generation | 3D layout generation · 3D-FRONT living rooms | OBA (m²): 341.1 -> 167.8 | Helped |
| Use DeBaRA instead of DiffuScene for bounded dining-room layout generation | 3D layout generation · 3D-FRONT dining rooms | FID: 23.06 -> 22.04 | Helped |
| Use DeBaRA instead of DiffuScene for bounded dining-room layout generation | 3D layout generation · 3D-FRONT dining rooms | OBA (m²): 266.4 -> 132.8 | Helped |
| Select LLM-generated semantics with SSE instead of selecting randomly | DeBaRA · 3D-FRONT living rooms scene synthesis | FID: 21.58 -> 20.59 | Helped |
| Select LLM-generated semantics with SSE instead of selecting randomly | DeBaRA · 3D-FRONT dining rooms scene synthesis | FID: 27.09 -> 24.50 | Helped |
| Add conditioning dropout (rate 0.2) to training (ablation) | DeBaRA with semantic-aware Chamfer distance · 3D-FRONT living rooms layout generation | FID: 19.89 -> 18.89 | Helped |
| Add conditioning dropout (rate 0.2) to training (ablation) | DeBaRA with semantic-aware Chamfer distance · 3D-FRONT dining rooms layout generation | FID: 22.60 -> 22.04 | Helped |
| Use semantic-aware Chamfer distance instead of MSE (ablation) | DeBaRA with dropout rate 0.0 · 3D-FRONT living rooms layout generation | FID: 21.66 -> 19.89 | Helped |
| Use semantic-aware Chamfer distance instead of standard Chamfer distance (ablation) | DeBaRA with dropout rate 0.0 · 3D-FRONT dining rooms layout generation | FID: 25.21 -> 22.60 | Helped |
| Replace 1000-step DDPM sampling with 25-step EDM sampling (ablation) | DeBaRA · 3D-FRONT living rooms layout generation | generation time (s): 5.144 -> 0.247 | Helped |
| Replace 1000-step DDPM sampling with 25-step EDM sampling (ablation) | DeBaRA · 3D-FRONT living rooms layout generation | FID: 21.12 -> 19.53 | Helped |
Surprising findings
- On the dining room set, DeBaRA's SCA (52.4%) is better, but DiffuScene obtains the best FID in the scene synthesis table (23.85 vs DeBaRA 24.50 with SSE), and DiffuScene has better SCA on living rooms in the scene synthesis table.
Limitations
- Does not enforce physical constraints, so objects can collide.
- Assumes object categories come from a finite predefined set.
- No style consistency enforced between objects.
Experimental setup
Datasets: 3D-FRONT 3D-FUTURE
Benchmarks: 3D-FRONT living rooms 3D-FRONT dining rooms
Metrics: FID KID SCA OBA OBR OBN
Baselines: ATISS DiffuScene LayoutGPT LEGO-Net
Builds on: EDM DiffuScene ATISS classifier-free guidance diffusion classifiers
Closest papers by meaning
- DiffuScene: Denoising Diffusion Models for Generative Indoor Scene Synthesis (CVPR 2024 · similarity 0.90)
- Direct Numerical Layout Generation for 3D Indoor Scene Synthesis via Spatial Reasoning (NeurIPS 2025 · similarity 0.90)
- OptiScene: LLM-driven Indoor Scene Layout Generation via Scaled Human-aligned Data Synthesis and Multi-Stage Preference Optimization (NeurIPS 2025 · similarity 0.90)
- CASAGPT: Cuboid Arrangement and Scene Assembly for Interior Design (CVPR 2025 · similarity 0.89)
- SceneCraft: Layout-Guided 3D Scene Generation (NeurIPS 2024 · similarity 0.88)
- Hierarchically-Structured Open-Vocabulary Indoor Scene Synthesis with Pre-trained Large Language Model (AAAI 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.