This post decribes ArchiGAN, an adversarial deep learning network based on Pix2Pix to generate apartment and building floorplans.
Overview
ArchiGAN is a framework that leverages Generative Adversarial Networks (GANs) to assist architects in designing apartment buildings. The article discusses a three-step generation stack for creating floor plans, including building footprint massing, program repartition, and furniture layout, while emphasizing the potential of AI in architectural design.
What You'll Learn
1
How to implement a generative design pipeline using GANs for architecture
2
Why using Pix2Pix GAN models can enhance architectural design processes
3
When to apply statistical approaches in architectural design using AI
Prerequisites & Requirements
- Understanding of Generative Adversarial Networks (GANs)
- Familiarity with TensorFlow and Google Cloud Platform
Key Questions Answered
How does the ArchiGAN framework assist in apartment building design?
The ArchiGAN framework utilizes a three-step generation stack that includes footprint massing, program repartition, and furniture layout. By employing Pix2Pix GAN models, it allows architects to create complex designs while incorporating user input at each stage, enhancing the design process through AI.
What are the limitations of using GANs in architectural design?
Current limitations include the inability to ensure continuity of load-bearing walls across multiple floors and the challenge of transforming raster images into vector formats suitable for architectural tools. Addressing these issues is crucial for integrating GAN outputs into standard design practices.
What specific tasks do the models in the ArchiGAN framework perform?
The models in the ArchiGAN framework perform three distinct tasks: Model I generates building footprints based on parcel shapes, Model II handles program repartition and fenestration, and Model III focuses on furniture layout within the generated spaces, allowing for a comprehensive design process.
Why is user input important in the ArchiGAN design process?
User input is essential in the ArchiGAN design process as it allows architects to modify and refine the generated outputs at each stage. This interaction ensures that the final designs meet specific requirements and maintain architectural integrity while leveraging AI capabilities.
Key Statistics & Figures
Training time reduction
From over a day and a half to under 2 hours
This reduction was achieved by utilizing an NVIDIA Tesla V100 GPU on Google Cloud Platform for training the GAN models.
Number of apartment plans used for training Model II
800+ plans
These plans were properly annotated and provided in pairs to train the network for program repartition and fenestration.
Technologies & Tools
Some links below are affiliate links. We may earn a commission if you make a purchase.
Machine Learning Framework
Pix2pix
Used for image-to-image translation in the design process.
Machine Learning Framework
Tensorflow
Utilized for implementing the GAN models and training processes.
Cloud Computing
Google Cloud Platform
Provided the infrastructure for training the models using NVIDIA GPUs.
Key Actionable Insights
1Integrate user feedback in each step of the design process to enhance the quality of architectural outputs.User input allows for iterative refinement, ensuring that the generated designs align with real-world requirements and preferences, ultimately leading to better architectural solutions.
2Utilize the ArchiGAN framework to explore innovative design possibilities that traditional methods may overlook.By leveraging AI and GANs, architects can experiment with unconventional designs and layouts, pushing the boundaries of creativity in architectural practice.
3Consider the scalability of the GAN models when designing multi-apartment buildings.The ability to process multiple units simultaneously can significantly streamline the design process for larger projects, making it essential for architects working on urban developments.
Common Pitfalls
1
Failing to ensure continuity of load-bearing walls across multiple floors can lead to structural issues.
This occurs because the internal structure is laid out differently for each unit, which can misalign load-bearing elements. Architects should consider specifying load-bearing positions in the input to address this.
2
Not transforming raster outputs into vector formats can hinder practical application in architectural design.
Since GAN outputs are pixel-based, they cannot be directly used by architects. A conversion process is necessary for integration with standard design tools.
Related Concepts
Generative Adversarial Networks (gans)
Architectural Design Processes
Machine Learning In Architecture
Ai-assisted Design Tools