Research map / Graph learning / Graph self-supervised learning
GSDiff: Synthesizing Vector Floorplans via Geometry-enhanced Structural Graph 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 GSDiff, a diffusion-plus-Transformer framework that generates vector floorplans as structural graphs (wall junctions via diffusion, wall segments via edge prediction) with an alignment loss and an edge-interpolation self-supervision to improve geometric consistency.
Problem
Automatic vector floorplan generation suffers from geometric inconsistencies (misalignment, overlaps, gaps), heavy post-processing, and limited constrained generation in existing rule-based and learning-based methods.
Approach
Decompose floorplan synthesis into structural graph generation: a diffusion Transformer generates wall junction nodes with a multi-base (real, binary, quaternary, octal, hexadecimal) alignment loss, and a Transformer edge predictor is trained with a random-interpolation coefficient self-supervision task; rooms are extracted as minimal polygon loops.
Key results
- On topology-constrained generation, the 125M model reaches FID 6.82 and KID 1.79 versus HouseDiffusion's 11.87 and 7.23, with GED 0.49 vs 2.59.
- On boundary-constrained generation, FID 7.50 and KID 0.56 vs WallPlan's 9.07 and 1.02 and Graph2Plan's 8.40 and 1.34.
- Ablation on RPLAN unconstrained generation: full Node+Edge reaches FID 4.83 and FE 0.95% vs 6.23 and 1.35% without enhancements.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GSDiff instead of HouseDiffusion | Topology-constrained floorplan generation · RPLAN topology-constrained | FID: 11.87 -> 6.82 | Helped |
| use GSDiff instead of HouseDiffusion | Topology-constrained floorplan generation · RPLAN topology-constrained | GED: 2.59 -> 0.49 | Helped |
| use GSDiff instead of WallPlan | Boundary-constrained floorplan generation · RPLAN boundary-constrained | FID: 9.07 -> 7.50 | Helped |
| remove edge perception enhancement (ablation) | GSDiff · GSDiff ablation | FID: 4.83 -> 4.96 | Hurt |
| remove edge perception enhancement (ablation) | GSDiff · GSDiff ablation | FE (%): 0.95 -> 1.15 | Hurt |
| remove node alignment enhancement (ablation) | GSDiff · GSDiff ablation | FID: 4.83 -> 5.72 | Hurt |
| remove node alignment enhancement (ablation) | GSDiff · GSDiff ablation | FE (%): 0.95 -> 1.13 | Hurt |
| replace mixed-base node alignment enhancement with real-valued-only alignment (ablation) | GSDiff · GSDiff ablation | FID: 4.83 -> 5.84 | Hurt |
| replace mixed-base node alignment enhancement with binary-only alignment (ablation) | GSDiff · GSDiff ablation | FID: 4.83 -> 5.84 | Hurt |
| add real-valued-only node alignment enhancement without edge perception enhancement (ablation) | GSDiff · GSDiff ablation | FID: 6.23 -> 6.18 | Helped |
| add binary-only node alignment enhancement without edge perception enhancement (ablation) | GSDiff · GSDiff ablation | FID: 6.23 -> 5.90 | Helped |
| add mixed-base node alignment enhancement without edge perception enhancement (ablation) | GSDiff · GSDiff ablation | FID: 6.23 -> 4.96 | Helped |
Limitations
- Depends on high-quality training data for learning node and edge dependencies
- Scalability challenges for larger projects
- Limited to constraint types it handles (boundary and topology); LIFULL results limited by annotation quality
Experimental setup
Datasets: RPLAN LIFULL
Benchmarks: RPLAN LIFULL
Metrics: FID KID GED FE AE room count ratio to ground truth
Baselines: Graph2Plan WallPlan House-GAN++ HouseDiffusion
Builds on: HouseDiffusion Graph2Plan WallPlan
Closest papers by meaning
- Constrained Layout Generation with Factor Graphs (CVPR 2024 · similarity 0.87)
- MaskPLAN: Masked Generative Layout Planning from Partial Input (CVPR 2024 · similarity 0.85)
- OptiScene: LLM-driven Indoor Scene Layout Generation via Scaled Human-aligned Data Synthesis and Multi-Stage Preference Optimization (NeurIPS 2025 · similarity 0.85)
- CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction (NeurIPS 2025 · similarity 0.84)
- HouseLayout3D: A Benchmark and Training-free Baseline for 3D Layout Estimation in the Wild (NeurIPS 2025 · similarity 0.84)
- Optimize Any Topology: A Foundation Model for Shape- and Resolution-Free Structural Topology Optimization (NeurIPS 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.