Research map / 3D vision / 3D generation
CASAGPT: Cuboid Arrangement and Scene Assembly for Interior Design
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 CASAGPT, a Llama-3-style autoregressive transformer that generates indoor scenes as sequences of cuboid primitives, with rejection sampling to remove collisions, and releases a cleaned dataset 3DFRONT-NC.
Problem
Indoor scene synthesis generators produce physically implausible scenes with object collisions, largely because bounding-box representations do not match actual mesh intersections and training data (3D-FRONT) contains intersecting objects.
Approach
Decompose each furniture mesh into cuboid primitives and model scenes as autoregressive token sequences of these cuboids, then fine-tune with rejection sampling that discards samples with cuboid intersections; also refine the training data via cuboid-based intersection avoidance.
Key results
- Achieves 90.6% non-intersection rate (NIRate) for bedrooms on 3DFRONT-NC, versus 79.1% for DiffuScene and 80.2% for ATISS on the same dataset.
- Bedroom FID 24.7 on 3DFRONT-NC versus 27.0 for ATISS; CIoU 0.77 versus 1.15 for ATISS.
- Rejection sampling iterations steadily improve CIoU and NIRate (bedroom NIRate 80.3 -> 87.0 -> 90.6), while FID slightly worsens in later iterations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CASAGPT instead of DiffuScene | Scene generation on 3D-FRONT bedrooms · Bedroom, original dataset | NIRate: 79.1 -> 80.6 | Helped |
| Use CASAGPT instead of DiffuScene | Scene generation on 3DFRONT-NC bedrooms · Bedroom, 3DFRONT-NC | NIRate: 79.1 -> 90.6 | Helped |
| Use CASAGPT instead of DiffuScene | Scene generation on 3DFRONT-NC living rooms · Living room, 3DFRONT-NC | NIRate: 65.8 -> 89.0 | Helped |
| Use cuboid rather than bounding-box representation for dataset refinement (ablation) | 3D-FRONT dataset refinement · Bedroom | NIRate: 96.3 -> 99.1 | Helped |
| Use cuboid rather than bounding-box representation for dataset refinement (ablation) | 3D-FRONT dataset refinement · Living room | FID: 19.0 -> 16.3 | Helped |
| Use cuboid instead of bounding-box sequence modeling (ablation) | CASAGPT · Bedroom | NIRate: 71.8 -> 76.8 | Helped |
| Use cuboid instead of bounding-box sequence modeling (ablation) | CASAGPT · Living room | FID: 38.9 -> 30.3 | Helped |
| Add rejection sampling through iteration 1 (ablation) | CASAGPT cuboid sequence model · Bedroom | NIRate: 76.8 -> 80.3 | Helped |
| Add rejection sampling through iteration 2 (ablation) | CASAGPT cuboid sequence model · Bedroom | NIRate: 80.3 -> 87.0 | Helped |
| Add rejection sampling through iteration 3 (ablation) | CASAGPT cuboid sequence model · Bedroom | NIRate: 87.0 -> 90.6 | Helped |
| Add rejection sampling through iteration 3 (ablation) | CASAGPT cuboid sequence model · Living room | NIRate: 72.5 -> 89.0 | Helped |
| Continue rejection sampling through iteration 3 (ablation) | CASAGPT cuboid sequence model · Bedroom | FID: 24.2 -> 24.7 | Hurt |
Surprising findings
- Bounding-box-based dataset refinement reduces intersections but disturbs non-intersecting table-chair groups, whereas cuboid-based refinement preserves them.
- RoPE underperformed learned position encoding on this short-sequence task.
Limitations
- In complex scenes with limited floor space, some objects still shift off the floor plane.
- FID slightly degrades in later rejection-sampling iterations as the distribution diverges from the original dataset.
- LayoutGPT only handles bedroom and living room scenes, so CIoU and NIRate are not applicable to it on all datasets.
Experimental setup
Datasets: 3D-FRONT 3D-FUTURE 3DFRONT-NC
Benchmarks: 3D-FRONT 3DFRONT-NC
Metrics: FID CKL CIoU NIRate non-occlusion rate
Baselines: LayoutGPT ATISS DiffuScene
Builds on: ATISS Llama-3 rejection sampling LayoutGPT
Closest papers by meaning
- DeBaRA: Denoising-Based 3D Room Arrangement Generation (NeurIPS 2024 · similarity 0.89)
- Direct Numerical Layout Generation for 3D Indoor Scene Synthesis via Spatial Reasoning (NeurIPS 2025 · similarity 0.88)
- OptiScene: LLM-driven Indoor Scene Layout Generation via Scaled Human-aligned Data Synthesis and Multi-Stage Preference Optimization (NeurIPS 2025 · similarity 0.87)
- Hierarchically-Structured Open-Vocabulary Indoor Scene Synthesis with Pre-trained Large Language Model (AAAI 2025 · similarity 0.87)
- InternScenes: A Large-scale Simulatable Indoor Scene Dataset with Realistic Layouts (NeurIPS 2025 · similarity 0.87)
- FreeScene: Mixed Graph Diffusion for 3D Scene Synthesis from Free Prompts (CVPR 2025 · 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.