Research map / Graph learning / Knowledge and temporal graphs
Constrained Layout Generation with Factor Graphs
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 FP-FGNN, a factor graph neural network that represents each room with four bounding-box variables and higher-order factors to predict floorplan layouts satisfying spatial constraints.
Problem
Object-centric layout generation under spatial constraints (e.g., floorplan design), where single-node object representations fail to capture fine-grained and higher-order interactions between objects and boundaries.
Approach
Represent each room by four latent variables (xmin, xmax, ymin, ymax) and connect them via factor nodes for box geometry, typed spatial relations, boundary corner points, and a complete factor, then apply learnable bipartite message passing (a factor graph neural network) to predict box coordinates before rendering the layout image with a CRN.
Key results
- Box-level IOU-Macro rises from 0.6796 (Graph2Plan) to 0.8685 and IOU-Micro from 0.7402 to 0.9165 on RPLAN.
- FID improves from 21.29 (RPLAN) and 29.26 (Graph2Plan) to 12.87 for FP-FGNN generations.
- With partial information, relation accuracy for specified constraints stays around 94.7-96.6% while Graph2Plan drops to about 78%; inference takes 0.023 s.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FP-FGNN instead of Graph2Plan for bounding-box prediction | Floorplan generation on RPLAN · RPLAN test | Box-level IOU-Macro: 0.6796 -> 0.8685 | Helped |
| Use FP-FGNN instead of Graph2Plan for bounding-box prediction | Floorplan generation on RPLAN · RPLAN test | Box-level IOU-Micro: 0.7402 -> 0.9165 | Helped |
| Use FP-FGNN instead of Graph2Plan for bounding-box prediction | Floorplan generation on RPLAN · RPLAN test | Box-level accuracy: 0.8683 -> 0.8901 | Helped |
| Use FP-FGNN instead of the strongest baseline for pixel-level IOU-Macro | Floorplan generation on RPLAN · RPLAN test | Pixel-level IOU-Macro: 0.4812 -> 0.5954 | Helped |
| Use FP-FGNN instead of Graph2Plan for pixel-level prediction | Floorplan generation on RPLAN · RPLAN test | Pixel-level IOU-Micro: 0.7575 -> 0.8019 | Helped |
| Use FP-FGNN instead of Graph2Plan for constraint preservation | Floorplan generation on RPLAN · RPLAN test | Relation accuracy: 0.9358 -> 0.9680 | Helped |
| Use FP-FGNN instead of Graph2Plan for constraint preservation | Floorplan generation on RPLAN · RPLAN test | Location accuracy: 0.9805 -> 0.9913 | Helped |
| Use FP-FGNN in the generation pipeline instead of the strongest baseline | Floorplan generation · Floorplan generation | FID: 21.29 -> 12.87 | Helped |
Surprising findings
- Pixel-level IOU-Macro is lower than Graph2Plan's Micro score for all models because small classes like wall-in and entrance are hard to represent with boxes.
Limitations
- Two-stage pipeline relies on a separate CRN to render images from predicted boxes.
- Generative diversity requires an external Graph2Plan-style retrieval step; the model itself is predictive.
- Pixel-level macro IOU is lower than baselines due to small classes.
Experimental setup
Datasets: RPLAN Graph2Plan room-adjacency graph version of RPLAN
Benchmarks: RPLAN (Graph2Plan split)
Metrics: box-level IOU macro box-level IOU micro pixel-level IOU pixel accuracy relation accuracy location accuracy FID
Baselines: Graph2Plan HPGM RPLAN Constr-LGen
Builds on: Graph2Plan factor graph neural network CRN (cascaded refinement network)
Closest papers by meaning
- GSDiff: Synthesizing Vector Floorplans via Geometry-enhanced Structural Graph Generation (AAAI 2025 · similarity 0.87)
- OptiScene: LLM-driven Indoor Scene Layout Generation via Scaled Human-aligned Data Synthesis and Multi-Stage Preference Optimization (NeurIPS 2025 · similarity 0.86)
- MaskPLAN: Masked Generative Layout Planning from Partial Input (CVPR 2024 · similarity 0.86)
- Direct Numerical Layout Generation for 3D Indoor Scene Synthesis via Spatial Reasoning (NeurIPS 2025 · similarity 0.85)
- See it to Place it: Evolving Macro Placements with Vision Language Models (NeurIPS 2026 · similarity 0.84)
- DeBaRA: Denoising-Based 3D Room Arrangement Generation (NeurIPS 2024 · 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.