SHOWCASE_GUIDE.md

FloorGen: Publication, Video Showcase & Viral Demo Playbook

This document details how to take FloorGen from code into a publishable research paper, viral video demo, and interactive portfolio project.

---

1. What is the Core Use Case & Value Proposition?

🏢 Practical & Commercial Impact

1. Instant Architectural Drafting: Architects & designers spend hours drafting initial CAD bubble diagrams. FloorGen generates vector floorplans in under 500 milliseconds.
2. Real Estate & PropTech Generation: Generate hundreds of customized variations of apartments matching a building site boundary and client preferences.
3. Gaming & Virtual Worlds: Procedural generation of indoor building layouts for Unreal Engine 5 / Unity, robotics simulators (Habitat, AI2-THOR).

🔬 Academic Novelty (Why it's publishable)

• Problem: Standard generative models either hallucinate invalid rooms (LLMs) or produce blurry raster pixels (Stable Diffusion).
• FloorGen's Contribution: Retrieval-Augmented Vector Diffusion (RAG-Diffusion) that retrieves the top-$k$ nearest architectural exemplars from a dual vector (FAISS) + topological graph (Neo4j) index and injects them as continuous cross-attention priors into coordinate diffusion denoising.

---

2. Generating Eye-Catching Videos & Visuals

A. Automatic Diffusion Denoising Video / GIF

Run the animation generator to create a high-framerate animation showing Gaussian noise resolving into an architectural floorplan:

python -m floorgen.animate_diffusion --output diffusion_synthesis.gif --frames 30

This produces an animated demonstration of the reverse diffusion trajectory, showing the gradual crystallization of walls, doors, and room labels.

B. Interactive 3D Isometric View & Real-Time Scrubbing

Launch the Web UI:

python -m floorgen.demo.web_app --port 8000

• Open http://localhost:8000
• Click "3D Isometric Model" to see live 3D wall extrusions and lighting rendered in WebGL/Three.js with an orbiting camera.
• Drag the "Diffusion Reverse Trajectory" slider to scrub through diffusion timesteps $T=100 \to 0$ in real time.

---

3. How to Make It Publishable (Step-by-Step)

Step 1: Benchmark Submission & LaTeX Paper

• Use the comparative benchmark table in README.md containing FID, KID, Graph Edit Distance (GED), and Realism scores.
• Write a 4 to 8-page workshop/conference paper (targeted for CVPR / ICCV / NeurIPS Machine Learning for Design Workshop / SIGGRAPH or arXiv).
• Paper title template: "FloorGen: Retrieval-Augmented Generative Floorplan Synthesis via Structural Graph Diffusion".

Step 2: Hugging Face Spaces 1-Click Interactive Demo

• Upload the repository to Hugging Face Spaces (Gradio / HTML) under a free CPU instance.
• Tag with #GenerativeAI, #DiffusionModels, #Architecture, #RAG.

Step 3: Social & Video Showcase (LinkedIn / X / YouTube)

1. Hook (0:00 - 0:03): Screen recording of dragging the slider to watch the 3D floorplan extrude and materialize from noise.
2. Concept (0:03 - 0:15): Show the RAG retrieval module picking 5 similar real floorplans from FAISS.
3. Comparison (0:15 - 0:30): Side-by-side showing standard GAN baseline (messy overlaps) vs FloorGen RAG-Diffusion (crisp, regularized vector layout with doors).
4. Call to Action: Link to the GitHub repo and free Google Colab notebook.