GRE-Diff: Gaussian Room Embeddings for Structured Layout Diffusion¶
Conference: ECCV 2026
Paper: ECCV Official Link
Area: 3D Vision
Keywords: Floor Plan Generation, Layout Diffusion, Gaussian Room Embedding, Controllable Editing, Structural Design
TL;DR¶
GRE-Diff models individual rooms as continuous 2D Gaussian distributions (Gaussian Room Embeddings, GRE) to guide a vector diffusion process, achieving an FID of 4.36 and a 99.21% F1BC-RC on RPLAN while enabling robust multi-modal and GUI-based interactive floor plan editing.
Background & Motivation¶
Designing functional and aesthetically coherent apartment floor plans is an intrinsically difficult architectural task. Human designers must repeatedly balance external building boundary shapes, functional zoning, daylight/ventilation requirements, and structural constraints across an enormous combinatorial design space of topological and geometric room arrangements. While data-driven deep generative models have emerged to assist this process, existing formulations typically simplify floor plan generation into raster image painting, discrete semantic segmentation masks, or ordered sequences of polygon vertices.
These discrete formulations introduce three core challenges that undermine robust design automation. First, polygon vertex representations are highly fragile; they exhibit extreme sensitivity to clockwise/counter-clockwise ordering and starting vertex indexing, where small coordinate perturbations can easily collapse geometric validity. Second, achieving controllable generation under strict architectural boundaries while preserving semantic room functional intent remains difficult, often resulting in disjoint rooms or invalid exterior penetrations. Third, interactive editing suffers from uncontrolled structural error propagation: modifying an individual room (such as expanding a bedroom or relocating a bathroom) causes cascading geometric distortions in neighboring spaces, producing unassigned voids or global layout failure.
Drawing inspiration from continuous probabilistic spatial embeddings and controllable generative representations, the authors argue that the fundamental bottleneck lies in how room geometry is represented inside the generative process. The core idea is to represent each room as a continuous, isotropic 2D Gaussian distribution (Gaussian Room Embedding, GRE) that summarizes its centroid and spatial footprint, using this probabilistic prior to initialize and guide diffusion denoising for vectorized polygon synthesis and structure-preserving local editing.
Method¶
Overall Architecture¶
GRE-Diff operates across three coordinated stages: initial constraint specification, layout generation, and dual-mode interactive editing. User inputs consist of natural-language design prompts (parsed by an LLM such as Kimi-K2 into structured counts of functional room types) and exterior apartment boundary contours (represented as 64 uniformly sampled boundary points). The generative engine pairs GuidanceNet for continuous spatial prior prediction with DenoisingNet for vector polygon layout decoding.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
In["User Input<br/>Text Prompt / Boundary Contour"] --> Spec["LLM Constraint Parsing<br/>Extract Room Types & Boundary Points"]
Spec --> DPEnc["Dual-Path Encoder<br/>Decoupled Semantic/Polygon & Local Boundary Attention"]
DPEnc --> AR["Autoregressive Transformer<br/>Sequential Room Gaussian Prediction (ฮผ_i, ฯ_i)"]
AR --> Samp["Gaussian Latent State Sampling<br/>x_T^i ~ N(ฮผ_i, ฯ_i^2 I)"]
Samp --> Denoise["Dual-Attention DenoisingNet<br/>Cross-Boundary Conditioning + Inter-Room Self-Attention"]
Denoise --> PolyOut["Vector Polygon Output<br/>Closed Room Layouts"]
PolyOut -.->|Re-encoded as P_token| Edit["Interactive Editing<br/>Add / Delete / Move / Anchor Hard-Copy"]
Edit -.-> DPEnc
Key Designs¶
1. Gaussian Room Embedding: Continuous Probabilistic Prior Representation
To resolve the vulnerability of discrete polygon vertex sequences to coordinate jitter and starting-vertex permutation, the model represents each room \(i\) as a continuous 2D isotropic Gaussian distribution \(g_i = \mathcal{N}(\mu_i, \sigma_i^2 I)\). Here, the centroid \(\mu_i \in \mathbb{R}^2\) captures global room placement, while the scalar scale \(\sigma_i\) approximates the room's coarse spatial footprint and neighborhood influence. Instead of replacing explicit polygon vectors \(x_0^i = [v_1^i, \dots, v_N^i]\), the GRE functions as an intermediate latent state. The diffusion trajectory is directly initialized from this continuous distribution as \(x_T^i \sim \mathcal{N}(\mu_i, \sigma_i^2 I)\), anchoring high-dimensional vertex generation within valid geometric envelopes and neutralizing the instability of ordered discrete sequences.
2. GuidanceNet: Decoupled Multi-Condition Encoding and Autoregressive Prediction
To unify multi-modal design intentions without mutual interference, GuidanceNet tokenizes inputs into semantic tokens \(S_{\text{token}}\), boundary contour tokens \(B_{\text{token}}\), and existing polygon tokens \(P_{\text{token}}\). The Dual-Path Encoder assigns independent global attention modules to semantic and polygon tokens, omitting cross-attention between them so functional semantics and geometric configurations are learned in separate feature subspaces before fusion. A parallel local attention pathway aligns boundary tokens along the building contour. A task adapter modulates polygon cues (\(\gamma=0\) for initial layout synthesis, \(\gamma=1\) for refinement). Conditioned on this fused representation, an autoregressive Transformer (ARTrans) sequentially generates Gaussian parameters \((\mu_i, \sigma_i)\) starting from an \(\langle\text{SOS}\rangle\) token, explicitly capturing room adjacency, functional grouping, and relative area budgeting.
3. Dual-Attention DenoisingNet and Anchor Hard-Constraint Editing
DenoisingNet decodes the continuous Gaussian priors into explicit closed polygonal geometries. It incorporates a boundary encoder that injects architectural perimeter constraints via cross-attention, alongside a self-attention pathway that models topological adjacency and repulsion between distinct room polygons. During interactive design, GRE-Diff provides Add, Delete, Move, Repurpose, and Anchor operations. For localized editing, rather than regenerating the full layout, user-anchored rooms are preserved via hard-constraint index masking: after each reverse diffusion step, the anchored polygon vertices are copied verbatim from the initial design back into the state buffer. Only editable rooms are resampled under the updated GRE priors and residual boundary conditions, eliminating cascading drift and guaranteeing non-destructive local modification.
Loss & Training¶
The framework is trained for 300 epochs across 8 NVIDIA Quadro GV100 GPUs with a per-GPU batch size of 40 using AdamW (initial learning rate \(2 \times 10^{-4}\), linear warm-up, and cosine learning rate decay). The reverse diffusion step follows: $\(x_{t-1}^i = \frac{1}{\sqrt{\alpha_t}}\left( x_t^i - \frac{1-\alpha_t}{\sqrt{1-\bar{\alpha}_t}}\mathcal{D}(x_t, i, y) \right) + \sigma_t z^i\)$ where \(\mathcal{D}(x_t, i, y)\) predicts the noise residuals conditioned on structural guidance \(y\) (apartment contour and anchor masks), steering sampled room coordinates into sharp, non-overlapping polygonal boundaries.
Key Experimental Results¶
Main Results¶
On the RPLAN benchmark (72,709 training floor plans, 8,079 test plans spanning 6 room categories), performance was evaluated over 5 runs across 512 randomly sampled test layouts against 7 state-of-the-art baselines (data transcribed from Table 1):
| Method | Boundary | Bubble | Room Type | FID (โ) | KID (ร10โปยณ, โ) | MMD (โ) | COV (%, โ) | BC (%, โ) | RC (%, โ) | F1_BC-RC (%, โ) | Params (M, โ) | Inference Time (s, โ) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Graph2Plan | โ | โ | โ | 4.81 | 1.65 | 0.124 | 94.92 | 100.00 | 83.11 | 90.78 | 8 | 0.41 |
| iPLAN | โ | โ | โ | 12.63 | 12.20 | 0.148 | 75.11 | 47.53 | 85.65 | 60.75 | 46 | 1.61 |
| HouseDiffusion | โ | โ | โ | 6.55 | 2.83 | 0.193 | 76.36 | โ | 93.35 | โ | 27 | 8.76 |
| WallPLAN | โ | โ | โ | 6.73 | 4.46 | 0.149 | 58.40 | 99.01 | โ | โ | 106 | 0.72 |
| MaskPLAN | โ | โ | โ | 23.54 | 22.84 | 0.190 | 67.18 | 100.00 | 20.12 | 33.50 | 998 | 2.68 |
| GSDiff | โ | โ | โ | 4.55 | 1.13 | 0.118 | 86.13 | 91.61 | โ | โ | 137 | 0.67 |
| ChatHouseDiffusion | โ | โ | โ | 5.21 | 1.31 | 0.145 | 83.01 | 99.80 | 53.91 | 70.01 | 82 | 1.15 |
| GRE-Diff (Ours) | โ | โ | โ | 4.36 | 0.96 | 0.107 | 96.09 | 98.44 | 100.00 | 99.21 | 60 | 0.91 |
Ablation Study¶
Ablation on Global Attention (GA) and Local Attention (LA) inside GuidanceNet (data transcribed from Table 3):
| GA | LA | FID (โ) | KID (ร10โปยณ, โ) | MMD (โ) | COV (%, โ) | BC (%, โ) | RC (%, โ) | F1_BC-RC (%, โ) |
|---|---|---|---|---|---|---|---|---|
| โ | โ | 4.57 | 0.87 | 0.153 | 86.52 | 96.67 | 99.80 | 98.21 |
| โ | โ | 4.49 | 0.82 | 0.145 | 86.71 | 97.07 | 99.41 | 98.23 |
| โ (Full Model) | โ (Full Model) | 4.36 | 0.96 | 0.107 | 96.09 | 98.44 | 100.00 | 99.21 |
Editing performance across operation types (data transcribed from Table 2):
| Method | Add Tiny-ROE (%, โ) | Add F1 (%, โ) | Delete Tiny-ROE (%, โ) | Delete F1 (%, โ) | Anchor & Move Tiny-ROE (%, โ) | Anchor & Move F1 (%, โ) |
|---|---|---|---|---|---|---|
| MaskPLAN | 0.00 | 49.47 | 0.00 | 24.09 | 0.00 | 66.05 |
| ChatHouseDiffusion | 68.49 | 68.45 | 10.96 | 58.26 | 4.11 | 73.06 |
| GRE-Diff (Ours) | 0.00 | 86.27 | 0.00 | 98.58 | 0.00 | 97.59 |
Key Findings¶
- Superior Controllability without Bubble Graphs: GRE-Diff requires no input topological bubble graph, yet achieves a 100.00% room-type constraint satisfaction (RC) and a 98.44% boundary containment score (BC), leading to a 99.21% F1BC-RC that outperforms ChatHouseDiffusion by +29.20%.
- Zero-Error Editing Fidelity: In interactive editing benchmarks, GRE-Diff attains a 0.00% Tiny-ROE error rate across all operations, with F1 reaching 97.59% on Anchor & Move (+24.53% over ChatHouseDiffusion), demonstrating that the anchor hard-copy mechanism reliably protects untouched spaces.
- High Downstream Usability: Across 19,680 generated samples, 95.56% of layouts are completely overlap-free, and 96.35% possess an unbroken door-connected accessibility graph forming a single connected component.
- User Preference: In a double-blind user study with 58 graduate students over 1,740 evaluation pairs, GRE-Diff layouts were preferred over ground truth in 60.69% of cases and over Graph2Plan in 70.12% of cases.
Highlights & Insights¶
- Probabilistic Gaussian Embeddings as a Spatial Bridge: Associating rooms with continuous 2D Gaussians effectively bridges abstract functional semantics and sharp polygon boundaries, bypassing the permutation sensitivity inherent in discrete vertex indexing.
- Hard-Constraint Anchor Mechanism in Diffusion Inversion: Copying fixed polygon coordinates back into the latent state buffer after each denoising step guarantees zero degradation in anchored zones, providing an elegant solution to interactive generative editing.
- Broad Transferability to Constrained Layout Problems: The two-stage paradigm (autoregressive continuous spatial prior estimation followed by conditional geometry diffusion) is directly applicable to 3D indoor scene synthesis, architectural site planning, and VLSI floorplanning.
Limitations & Future Work¶
- Author-Acknowledged Limitations: The model struggles with highly non-rectilinear, non-Manhattan boundaries or extreme geometric aspect ratios, occasionally yielding uneven vertex spacing or distorted room geometries.
- Self-Identified Limitations: The isotropic formulation (\(\sigma_i^2 I\)) assumes circular influence regions, neglecting directional orientation and high aspect ratios characteristic of long hallways or narrow balconies. An anisotropic covariance formulation with explicit rotation parameters could better capture elongated rooms.
- Future Directions: Extending the framework to handle multi-floor layouts and integrating fine-grained furniture placement and functional object layout into a unified generative pipeline.
Related Work & Insights¶
- vs Graph2Plan: Graph2Plan relies on non-neural retrieval post-processing; removing this retrieval degrades its FID from 4.81 to 20.10 and drops RC to 50.19%. In contrast, GRE-Diff synthesizes vectorized polygons end-to-end without retrieval databases.
- vs HouseDiffusion / ChatHouseDiffusion: Prior vector diffusion approaches apply noise directly to polygon vertices without room-level spatial guidance, resulting in unassigned gaps and disconnected spaces. GRE-Diff injects GRE priors into diffusion initialization and trajectory updates, maintaining structural coherence.
- vs MaskPLAN: MaskPLAN relies on 998M parameters and raster mask prediction with heavy polygonization post-processing, often resulting in boundary drift; GRE-Diff uses a compact 60M parameter network to infer vector layouts directly in 0.91 seconds.
Rating¶
- Novelty: โญโญโญโญโ [Gaussian room embeddings provide a continuous probabilistic bridge for structured layout generation]
- Experimental Thoroughness: โญโญโญโญโญ [Extensive comparisons across 7 SOTAs, downstream topological metrics, editing ablations, and a 58-participant user study]
- Writing Quality: โญโญโญโญโญ [Clear structural narrative, disciplined mathematical formulation, and high-quality figures]
- Value: โญโญโญโญโญ [A practical and highly controllable blueprint for computer-aided architectural design and floor plan editing]