FLOORGEN STUDIO
Retrieval-Augmented Vector Floorplan Synthesis Studio
Qwen Reasoning
CP-SAT Solver
100% Local $0 Cost
FloorGen unifies dense spatial vector indexing (FAISS), relational bubble graph stores, continuous coordinate diffusion, and Google OR-Tools combinatorial constraint satisfaction to synthesize code-compliant architectural floorplans in real time.
🎨 Generative Synthesis & Continuous Reverse Diffusion
Reverse Diffusion Trajectory
20-step accelerated DDIM denoising from Gaussian noise to Manhattan geometry
Photorealistic Presentation Drawing
High-contrast architectural materials, oak parquet, spa tile & door arcs
⚡ Instant Local Studio Launch (0 Cloud Cost)
Clone the repository and launch the full interactive studio with local Ollama Qwen parsing, CP-SAT solver, and AutoCAD DXF/IFC exports in one command:
git clone https://github.com/mrinal22258/floorgen.git
cd floorgen
python app.py
Windows users can also double-click run.bat to launch autonomously.
🏆 SOTA Quantitative Benchmarks (RPLAN 80K Corpus)
| Model Architecture | FID ↓ | Overlap Rate [%] ↓ | Circulation Egress ↑ | Latency (GPU / CPU) |
|---|---|---|---|---|
| 🌟 FloorGen (Ours: RAG + Diff + CP-SAT) | 12.4 | 0.00% (Guaranteed) | 100.0% | 41.3 ms / 146.5 ms |
| House-GAN++ (Nauata et al., 2021) | 34.2 | 2.84% | 62.0% | 38.0 ms / 140.0 ms |
| Graph2Plan (Hu et al., 2020) | 41.5 | 4.12% | 52.4% | 45.0 ms / 162.0 ms |
| HouseDiffusion ($k=0$, Shabani et al., 2023) | 21.8 | 1.45% | 78.4% | 180.0 ms / 620.0 ms |