Space Syntax-guided Post-training for Residential Floor Plan Generation

This paper proposes Space Syntax-guided Post-training (SSPT), a framework that integrates architectural theory into residential floor plan generation via a non-differentiable oracle and reinforcement learning to enhance public space dominance and functional hierarchy while outperforming distribution-fitted baselines in efficiency and stability.

The Gist

Imagine you are teaching a robot chef to cook the perfect family dinner. You show the robot thousands of photos of real meals (the pre-training phase). The robot learns to chop vegetables and fry eggs perfectly. It knows how to make a dish that looks like a meal.

But here's the problem: The robot doesn't understand the logic of a meal. It might put the salt shaker in the middle of the table where everyone can reach it (good), but it might also put the trash can right next to the dining table, or the toilet in the center of the living room. It follows the visual patterns but misses the "common sense" of how a home should actually work.

This paper, "Space Syntax-guided Post-training for Residential Floor Plan Generation," is about teaching that robot chef the logic of a home after it has already learned to draw the walls.

Here is the breakdown in simple terms:

1. The Problem: The Robot Draws "Pretty" but "Wrong" Homes

Current AI models are great at drawing floor plans that look realistic. They know a bedroom usually has a bed and a bathroom has a toilet. However, they often mess up the flow.

• The Issue: In a real home, the living room is the "heart" of the house. It should be easy to get to from everywhere. The bedroom should be a quiet, private corner.
• The AI Mistake: The AI sometimes puts the living room in a dark corner or makes the hallway the most important room. It's like a chef who puts the dessert on the main plate and the steak on the side. It looks like food, but it's not right.

2. The Solution: The "Architectural GPS" (Space Syntax)

The authors introduce a tool called Space Syntax. Think of this as a smart GPS for social flow.

• Instead of just looking at the walls, this tool asks: "If I stand in the living room, how many steps does it take to get to the kitchen? How many steps to the bedroom?"
• It calculates a score called "Integration."
  • High Integration: A room is central and easy to reach (like a town square).
  • Low Integration: A room is hidden and hard to reach (like a secret basement).
• The Rule: In a good house, the Living Room should have the highest "Integration" score. The Bedroom should have a lower score.

3. The Magic Trick: "Post-Training" (The Coach)

The AI is already trained to draw houses. The authors didn't want to re-teach the AI from scratch. Instead, they added a Coach that watches the AI work and gives it a score.

They created two ways for the Coach to teach the AI:

Strategy A: The "Filter" (SSPT-Iter)

• How it works: The AI draws 1,000 houses. The Coach checks them all. It throws away the 900 bad ones (where the living room is in the wrong spot) and keeps the 100 best ones.
• The Lesson: The AI re-trains itself only on those 100 good houses.
• The Result: It learns the rule, but it's slow and expensive because it has to draw thousands of houses just to find a few good ones.

Strategy B: The "Reward" (SSPT-PPO)

• How it works: This is like a video game. The AI draws a house. The Coach immediately gives it a "score" based on the Space Syntax rules.
  • "Living room in the center? +10 points!"
  • "Hallway blocking the door? -5 points!"
• The Lesson: The AI learns to maximize its score. It adjusts its drawing style in real-time to get more points.
• The Result: This is 11 times faster than Strategy A and produces much more consistent, high-quality homes.

4. The "Exam" (SSPT-Bench)

To prove their method works, they created a special test called SSPT-Bench.

• They taught the AI on houses with 7 rooms or fewer.
• Then, they tested it on houses with 8 rooms (a situation the AI had never seen before).
• Why? To see if the AI truly learned the logic of a home, or if it just memorized the training pictures.
• The Verdict: The AI trained with the "Reward" method (Strategy B) passed the test with flying colors. It understood that no matter how big the house is, the living room needs to be the central hub.

Summary Analogy

• Old AI: A student who memorized the answers to a math test but doesn't understand why the answer is correct. If you change the numbers slightly, they fail.
• New AI (SSPT): A student who learned the logic of math. Even if you give them a new, harder problem, they can solve it because they understand the rules.

The Bottom Line:
This paper shows that by adding a simple "logic check" (Space Syntax) after the AI has learned to draw, we can teach it to design homes that aren't just pretty pictures, but actually feel like real, livable spaces where the living room is the heart of the home. And the best part? They found a way to do it 11 times faster than before.

Technical Summary

1. Problem Statement

Current deep learning-based generative models for residential floor plans (e.g., diffusion models, GANs) are typically optimized to fit large-scale data distributions (like the RPLAN dataset). While effective at generating geometrically valid layouts, they often fail to capture critical architectural priors regarding spatial logic. Specifically:

• Loss of Functional Hierarchy: Models often fail to enforce the "public-space dominance" principle, where living rooms and foyers should act as topological integration cores (highly accessible and central), while private spaces (bedrooms, bathrooms) should be more segregated.
• Evaluation Gaps: Existing benchmarks rely on geometric metrics (IoU, FID) or simple adjacency rules, lacking metrics to evaluate deep spatial configurational logic (centrality, accessibility, and hierarchy).
• Data Quality: Large-scale datasets often contain layouts with problematic spatial logic or mislabeled functions, which are rarely screened out before training.
• Optimization Limitations: Standard pre-training cannot easily incorporate non-differentiable architectural constraints (like space syntax integration) without complex re-engineering of the generative backbone.

2. Methodology

The authors propose SSPT (Space Syntax-guided Post-training), a framework that injects architectural theory into pre-trained diffusion models via a non-differentiable oracle.

A. The Space Syntax Oracle

A deterministic pipeline that converts generated floor plans into quantitative architectural metrics:

1. Representation: Converts RPLAN-style 4-channel masks into a Rectangle-Space Graph.
   • Decomposition: Uses a greedy maximal-rectangle algorithm to decompose room cores into convex atoms.
   • Graph Construction: Nodes are rectangles; edges represent intra-room proximity and inter-room connectivity via doors.
2. Metric Computation: Calculates Spatial Integration (based on Space Syntax theory) for each node.
   • Integration ($Int_i$): Derived from Mean Depth ($MD$) and Relative Asymmetry ($RA$). High integration indicates a space is topologically central and accessible.
3. Scoring: Aggregates node scores to room instances to compute:
   • Public-Space Dominance: Ensures living rooms have higher integration than private rooms.
   • Validity Gates: Filters out layouts with missing rooms, abnormal area shares, or inverted hierarchies.

B. Post-Training Strategies

SSPT utilizes two distinct strategies to align the generator with the oracle's rewards:

1. SSPT-Iter (Iterative Retraining):
   • Generates a large batch of candidates.
   • Filters them using the Oracle's scores (Top-K selection).
   • Fine-tunes the diffusion model on this curated subset using standard denoising objectives.
2. SSPT-PPO (Reinforcement Learning):
   • Treats the diffusion reverse process as a Markov Decision Process (MDP).
   • Uses Proximal Policy Optimization (PPO) to directly optimize the policy based on the terminal reward provided by the Oracle.
   • Employs a "short respaced" policy (focusing on the low-noise denoising steps) to improve computational efficiency.

C. Benchmark: SSPT-Bench (Eval-8)

To ensure rigorous evaluation, the authors introduce an Out-of-Distribution (OOD) benchmark:

• Training: Models are post-trained on conditions with $\le 7$ rooms.
• Evaluation: Models are tested on 8-room layouts (Eval-8) to test generalization and robustness.
• Metrics: Includes Public-Space Dominance ($public\_score$), Living-Room Advantage ($living\_adv$), and Profile Distance ($d_{profile}$) against real-world data distributions.

3. Key Contributions

1. Architectural Prior Formalization: Translates the qualitative concept of "public-space dominance" into a computable, scalable integration metric derived from Space Syntax theory.
2. Automated Oracle & Dataset Screening: Developed a mask-to-graph oracle that can screen 80,000+ real-world layouts, identifying and removing spatially irrational plans, and establishing empirical baselines for integration hierarchies.
3. Unified Benchmark (SSPT-Bench): Created a reproducible OOD evaluation protocol (Eval-8) with a unified metric suite to diagnose functional hierarchy failures in AI-generated plans.
4. Efficient Post-Training Framework: Demonstrated that PPO-based post-training significantly outperforms iterative retraining in both architectural alignment and computational efficiency.

4. Experimental Results

Experiments were conducted on the HouseDiffusion baseline using the SSPT-Bench protocol.

• Performance Improvement:
  • Public-Space Dominance: SSPT-PPO improved the median $public\_score$ from 0.182 (Iterative) to 0.224, significantly narrowing the gap with real-world data (0.522).
  • Stability: SSPT-PPO reduced the variance (IQR) of metrics by ~20–30%, indicating more consistent generation of high-quality hierarchies.
  • Profile Alignment: The distance to the real-world integration profile ($d_{profile}$) decreased by 41.3% compared to the baseline, showing SSPT-PPO better restores the natural hierarchy (e.g., reducing over-centralized entrances and over-integrated private rooms).
• Efficiency:
  • SSPT-PPO achieved a 11.1× speedup per iteration compared to SSPT-Iter (0.75 hours vs. 8.30 hours).
  • When normalized by improvement gain, PPO was roughly 10× more efficient in terms of compute resources required to achieve architectural alignment.
• Case Analysis:
  • Correction of Errors: The framework successfully corrected common failure modes, such as "over-centralized threshold spaces" (entrances becoming too central) and "flattened hierarchies" (private rooms being too accessible).

5. Significance and Future Work

• Bridging Theory and AI: SSPT provides a scalable pathway to integrate established architectural theories (Space Syntax) into data-driven generative design without retraining the entire backbone from scratch.
• Reliability: It offers a method to automatically screen and validate large architectural datasets, ensuring that AI models learn from spatially rational examples.
• Generalizability: The framework is backbone-agnostic; while demonstrated on diffusion models, the oracle and PPO strategy can be applied to other generative architectures (Transformers, Graph-based models).
• Future Directions: The authors suggest expanding the oracle to include multi-objective criteria (circulation efficiency, zoning), accelerating the oracle pipeline, and integrating human-in-the-loop feedback for deployable design tools.

In conclusion, the paper demonstrates that post-training with a non-differentiable architectural oracle is a superior strategy for injecting domain-specific logic into generative AI, with PPO offering the most efficient and effective route to achieving high-fidelity, architecturally rational floor plans.