FLOORVERSE is an innovative platform that leverages cutting-edge AI to automate and optimize the generation of residential floorplans. By utilizing a Conditional Variational Autoencoder (CVAE) model, meticulously trained on the extensive RPlan dataset (comprising over 80,000 images), FLOORVERSE produces functionally coherent and aesthetically pleasing layouts based on user-defined constraints. The system seamlessly integrates a Flask-based backend with an interactive frontend, providing users with real-time, personalized design experiences.
At the heart of FLOORVERSE is a Conditional Variational Autoencoder (CVAE), designed to process visual data from floorplan images in conjunction with user-defined design parameters. The architecture is structured into the following key components:
- Encoder: Utilizes convolutional layers to efficiently extract spatial features directly from grayscale floorplan images.
- Reparameterization Trick: Employs this technique to sample latent vectors in a differentiable manner, crucial for effective backpropagation and model training.
- Decoder: Reconstructs existing floorplans or generates entirely new ones based on the learned latent representations and the specific user conditions provided.
- Training & Optimization: The model is trained using a composite loss function that combines binary cross-entropy for image reconstruction and KL divergence to ensure a well-structured latent space, leading to robust and reliable performance.
The FLOORVERSE workflow follows a streamlined process:
- Input Preprocessing: Raw floorplan images undergo preprocessing steps such as resizing and normalization to prepare them for model input.
- Feature Encoding: The Encoder network processes the preprocessed visual data, while condition vectors representing user inputs are processed through a multi-layer perceptron to capture relevant design constraints.
- Latent Space Representation: Fully connected layers are used to generate a compressed latent space representation that encapsulates the essence of the input floorplan and user conditions.
- Floorplan Generation: The Decoder network takes the latent space representation and user constraints to reconstruct or generate new floorplans that are tailored to the specified requirements.
The performance of FLOORVERSE has been rigorously evaluated through a combination of quantitative metrics and qualitative analyses to ensure both accuracy and design quality.
- Reconstruction Loss: 8571.3582 - Measures how well the model reconstructs input floorplans, indicating the fidelity of the encoder-decoder process.
- Total Loss: 8654.9813 - Represents the overall loss during training, combining reconstruction loss and KL divergence to optimize model performance.
- KL Divergence: 83.6231 - Quantifies the divergence between the learned latent space distribution and a standard normal distribution, ensuring a well-organized and continuous latent space for generation.
- Latent Space Traversals: Demonstrated the ability to smoothly interpolate between diverse design archetypes - Showcases the model's capability to generate a spectrum of floorplan styles and configurations by navigating the learned latent space.
- Principal Component Analysis (PCA): Highlighted spatial coherence in generated layouts - Visualizes the latent space and confirms that generated layouts maintain spatial relationships and structural integrity, resulting in realistic and well-structured floorplans.
git clone https://github.com/your-repo/FLOORVERSE.git
cd FLOORVERSEpip install -r requirements.txtpython app.pyComprehensive documentation for FLOORVERSE is available documentation/ directory to help you understand the project in detail, from setup to advanced usage and development.
This project was a collaborative effort by a team of dedicated individuals, including:
FLOORVERSE is built upon the foundation of extensive research in AI-driven architectural design. We gratefully acknowledge the insights and advancements from prior works in generative adversarial networks, graph neural networks, and diffusion-based approaches within the field. We extend our sincere gratitude to the creators of the RPlan dataset for providing a robust and invaluable resource that enabled the training of our model.