Building-GAN: Graph-Conditioned Architectural Volumetric Design Generation

Kai-Hung Chang, Chin-Yi Cheng, Jieliang Luo, Shingo Murata, Mehdi Nourbakhsh, Yoshito Tsuji

Abstract

Volumetric design is the first and critical step for professional building design, where architects not only depict the rough 3D geometry of the building but also specify the programs to form a 2D layout on each floor. Though 2D layout generation for a single story has been widely studied, there is no developed method for multi-story buildings. This paper focuses on volumetric design generation conditioned on an input program graph. Instead of outputting dense 3D voxels, we propose a new 3D representation named voxel graph that is both compact and expressive for building geometries. Our generator is a cross-modal graph neural network that uses a pointer mechanism to connect the input program graph and the output voxel graph, and the whole pipeline is trained using the adversarial framework. The generated designs are evaluated qualitatively by a user study and quantitatively using three metrics: quality, diversity, and connectivity accuracy. We show that our model generates realistic 3D volumetric designs and outperforms previous methods and baselines.

Introduction

In professional architecture, "volumetric design" serves as the critical bridge between a client's functional requirements and the physical reality of a building site. Designers must translate a "program"—a list of required spaces like offices, restrooms, and elevators—into a 3D massing that adheres to local building codes, site constraints, and vertical alignment needs. While automated 2D floor plan generation has seen significant progress, 3D multi-story generation remains a complex challenge due to the high dimensionality of 3D space and the intricate vertical dependencies inherent in architectural design.

Methodology

Building-GAN addresses these challenges through a generative adversarial framework specifically designed for architectural volumetric massing. The system automates the transition from abstract functional requirements to complex 3D forms using two key innovations:

Voxel Graph Representation

Traditional 3D representations like dense voxels are often too memory-intensive and produce "jagged" geometries unsuitable for architectural precision. Building-GAN utilizes a voxel graph, representing the design space as a graph of non-uniform cuboids. This approach efficiently handles irregular site boundaries and ensures that generated rooms maintain clean, "Manhattan-world" edges.

Cross-Modal Generation

The model takes a Hierarchical Program Graph (or "bubble diagram") as input, which encodes room types, story levels, and functional connections. To map these abstract nodes to physical space, the generator employs a Cross-Modal Pointer Module. For each voxel in the output space, the model "points" to a corresponding node in the program graph, allowing for a dynamic number of rooms per floor.

The Building-GAN pipeline: from abstract program graphs and space partitions to generated volumetric designs and final architectural realizations.

Dataset and Training

To train the model, the researchers developed a synthetic dataset of 120,000 volumetric designs in collaboration with professional architects. These designs were generated using parametric models based on real-world commercial building rules, covering heights up to 11 stories and various site conditions. The training process utilizes a Dual-Level Discriminator that evaluates designs at both the building level (for overall 3D coherence) and the story level (for local floor plan logic).

Key Insights

  • Vertical Consistency: By explicitly modeling vertical circulation elements like stairs and elevators as connected nodes across story sub-graphs, the model successfully maintains vertical alignment—a notorious difficulty in automated 3D design.
  • Architectural Logic: Analysis of the model's attention layers reveals a "growth" pattern similar to human design processes, often starting with the building's core and iteratively refining surrounding spaces.

Results

Building-GAN was evaluated against a 3D-extended baseline of House-GAN. Quantitatively, it achieved a Fréchet Inception Distance (FID) of 0.0845, significantly outperforming the baseline's 17.60. Connectivity accuracy also improved from 0.403 to 0.569.

A detailed view of the generated volumetric design (voxel graph), showing the assignment of functional programs to 3D volumes.

In a qualitative user study involving 20 professional architects, Building-GAN's designs received a quality score of 0.85, closely approaching the ground-truth human designs (0.92). Case studies suggest that this framework could reduce the initial design exploration phase from weeks to just a few days.

Conclusion and Future Work

While Building-GAN represents a significant step forward, challenges remain in ensuring perfect graph connectivity and avoiding fragmented rooms. Future research will focus on incorporating "hard" constraints for connectivity and Floor Area Ratio (FAR), as well as extending the voxel graph to support more organic, non-cuboid geometries.

Summary updated

Abstract

Volumetric design is the first and critical step for professional building design, where architects not only depict the rough 3D geometry of the building but also specify the programs to form a 2D layout on each floor. Though 2D layout generation for a single story has been widely studied, there is no developed method for multi-story buildings. This paper focuses on volumetric design generation conditioned on an input program graph. Instead of outputting dense 3D voxels, we propose a new 3D representation named voxel graph that is both compact and expressive for building geometries. Our generator is a cross-modal graph neural network that uses a pointer mechanism to connect the input program graph and the output voxel graph, and the whole pipeline is trained using the adversarial framework. The generated designs are evaluated qualitatively by a user study and quantitatively using three metrics: quality, diversity, and connectivity accuracy. We show that our model generates realistic 3D volumetric designs and outperforms previous methods and baselines.
BibTeX
@inproceedings{chang2021building,
  author = {Kai-Hung Chang and Chin-Yi Cheng and Jieliang Luo and Shingo Murata and Mehdi Nourbakhsh and Yoshito Tsuji},
  title = {{Building-GAN: Graph-Conditioned Architectural Volumetric Design Generation}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2021},
  url = {https://openreview.net/forum?id=S7jkedrMJib},
}
Figure