Boundary Cross-Attention (BCA) Module

• Boundary Cross-Attention (BCA) Module is an attention-based mechanism designed to enforce strict geometric constraints by conditioning generative models on user-specified spatial boundaries.
• It integrates with diffusion models via cross-attention where boundary encodings are used as keys and values to modulate floorplan synthesis at multiple scales.
• Empirical evaluations show that BCA significantly enhances boundary compatibility while balancing the trade-off between realistic adherence and design diversity.

Boundary Cross-Attention (BCA) Module

Boundary Cross-Attention (BCA) is an architectural module introduced to enhance geometric fidelity in generative models for floorplan synthesis, specifically within the context of diffusion models. It is designed to enable strict conditioning on user-specified spatial boundaries—typically represented as polygons—that define the feasible region for architectural layouts. The BCA module addresses a crucial need in conditional generative design, where generated samples must not only be realistic but must also precisely adhere to provided geometric constraints. The development and formal evaluation of BCA are set forth in “Boundary-Constrained Diffusion Models for Floorplan Generation: Balancing Realism and Diversity” (Stoppani et al., 2 Feb 2026), which establishes its effectiveness in improving boundary adherence without unduly sacrificing design diversity.

1. Motivation and Conceptual Foundations

Conventional diffusion models for layout generation commonly employ conditioning mechanisms—such as concatenation or attention-based modulation—on high-level graph structures or room adjacency. However, these strategies often fail to guarantee compliance with arbitrary user-defined geometric boundaries, leading to generated layouts that overlap, violate, or ignore the prescribed floorplate. BCA is explicitly engineered to enforce geometric consistency by integrating spatial boundary information directly as a cross-attention target within the model’s generative architecture.

The key goals addressed by the BCA module are:

• Strict boundary adherence: Ensuring all generated layout elements are subsumed within the user-defined polygonal boundary.
• Geometric consistency under constraints: Interpolating spatial relationships, especially near irregular or non-rectilinear boundaries.
• Compatibility with other conditioning signals: Operating in conjunction with other graph- or program-based constraints, such as room adjacency or labels.

2. Formal Architecture of the BCA Module

The BCA module is instantiated as a cross-attention operation within the diffusion model’s U-Net backbone, typically at multiple scales and time steps. The central components are as follows:

• Query ($Q$): Features derived from the model’s intermediate activations during the denoising step—these represent latent spatial building blocks of the synthesized floorplan.
• Key ($K$) and Value ($V$): Embeddings associated with sampled boundary points or parameterizations of the boundary polygon (e.g., via polylines or splines).

Mathematically, the BCA module can be expressed as a conventional multihead attention operation: $$\text{Attention}(Q,K,V) = \operatorname{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V$$ where $Q$ is a tensor representing the latent representation at the current generation step, and $K,V$ encode boundary geometry. In implementation, boundary points are uniformly sampled or interpolated along the boundary polygon, then sinusoidally or projection-encoded and fed into the $K,V$ mapping layers. The module replaces or augments ordinary self-attention in the U-Net, thus propagating explicit geometric priors into every decoding layer.

3. Integration into the Diffusion Process

Within the floorplan generation model, BCA modules are interleaved throughout the neural architecture:

• At each denoising timestep, the BCA injects explicit information about the boundary at multiple feature levels, both coarse and fine.
• Coupled with classifier-free guidance (CFG), the sampling procedure can be tuned via a guidance scale parameter $\lambda$, which controls the strictness of adherence to the boundary constraint.
• The BCA continuously modulates the generative process, such that any drift away from the prescribed boundary induces an explicit correction at subsequent steps.

This integration ensures that the geometric context is not just a global constraint but is accessible at every spatial and semantic scale during synthesis.

4. Empirical Effects and Quantitative Evaluation

Empirical evaluation of the BCA module is conducted by measuring the “boundary compatibility” (BC) of generated samples—specifically, the proportion of layout elements contained within the prescribed boundary without overlap or violation. The introduction of BCA is found to significantly improve boundary adherence metrics compared to baselines that only utilize concatenated or global boundary features. In controlled ablations:

• Models incorporating BCA demonstrate higher BC values across diverse boundary shapes and sizes.
• Increased $\lambda$ (CFG guidance scale) enforces stricter geometric compliance, but induces a measured decrease in design diversity, as quantified by a Diversity Score (DS) computed over feature embeddings of generated layouts.
• Prolonged training without sufficient diversity regularization can lead to “diversity collapse,” wherein FID and BC may continue to improve (realism/boundary) even as DS declines (loss of variety), emphasizing the balancing act between strict boundary conditioning and generative variability.

5. Theoretical and Practical Implications

The BCA module provides a principled architectural interface for geometric conditioning in generative models:

• From a theoretical standpoint, it situates hard spatial constraints as first-class citizens in attention-based representation learning.
• Practically, it enables practitioner control over the trade-off between geometric fidelity and creative diversity via the inference-time tuning of guidance parameters.
• The deployment of BCA encourages a shift toward multi-objective training and evaluation, wherein boundary compatibility, perceptual realism, and diversity are all monitored and optimized jointly.

Notably, BCA is not a domain-specific mechanism: Any generative image or layout synthesis setting requiring strict spatial boundary adherence can, in principle, benefit from its explicit cross-attention mechanism.

6. Limitations and Future Directions

While the BCA module raises the standard of geometric consistency in boundary-conditioned generation, it also introduces several axes of trade-off:

• Excessively strict boundary conditioning (high $\lambda$) depresses generative diversity, leading to mode collapse if not counterbalanced with explicit diversity regularization.
• The reliance on “feature embeddings” for boundary representation assumes the chosen encoding is both expressive and compatible with the underlying geometry; poorly chosen embeddings may impede model expressivity.
• The method is currently evaluated on architectural floorplans; extensions to 3D volumetric generation, irregular domains, or other semantic boundary regimes will require further investigation.

Extensions may include:

• Incorporation of continuous boundary parameterizations or higher-order geometric signals (e.g., signed distance fields, normals).
• Adaptive guidance scaling that adjusts the $\lambda$ parameter dynamically based on diversity feedback during inference.
• Explicit multi-objective optimization frameworks that simultaneously maximize BC, FID, and DS throughout training epochs.

7. Significance in Generative Design and Beyond

The BCA module represents a significant advancement in controllable, constraint-driven generative modeling for spatial design. By centering boundary adherence as an attention-mediated process, it enables precise user control over geometric outcomes, unlocking applications in automated architecture, robotic navigation, and any domain where generated content must conform to strict spatial envelopes. Its introduction also foregrounds the critical need to balance realism, constraint satisfaction, and diversity—an interplay that now forms a central research axis in generative model evaluation and development (Stoppani et al., 2 Feb 2026).