Text-to-Image Conditional GAN-Based Floor Plan Generator

To succeed with autonomous design processes within the architectural building design industry, conditional generative models are essential. One significant challenge with large language-image (LLI) models - such as Stable Diffusion, Adobe Firefly, Dall-E, and Midjourney - is their limited contextual understanding within specific design domains. However, precise conditional generation of floor plan designs is needed to unlock the full potential of custom-trained generative models. A custom-trained StyleGAN3 model, combined with a custom-trained captioning model and a translation model, merges the text embedding space with the GAN's latent space to enable text-to-image generation for highly contextual floor plan designs. This generative machine learning application is highly relevant for the early design stage of the architectural design process, enriching conceptual design development with faster, more efficient, and contextually meaningful design suggestions. The proposed framework contributes with a custom-trained text-to-floor plan generator utilizing a fine-tuned bootstrapped language-image pre-training (BLIP) model. Additionally, a comparison with a baseline model using a pre-trained BLIP model demonstrates that the fine-tuned text-to-image generator significantly outperforms the baseline model in accuracy of floor plan generation.