Project Overview
Project Summary:
Our "Architectural Drawing Generation with Recycled Materials" project aims to improve the productivity of architectural design while solving the construction waste problem. It utilizes generative AI to predict the required amount of building materials and promote the reuse of recycled materials. To generate architectural drawings, we leverage a multimodal input-enabled LCM (Latent Consistency Models) model based on Stable Diffusion. This model generates design drawings and birds-eye-views. Also we can predict the required amount of building materials based on the drawings.
Identifying the Challenge
The Social Problem:
Construction waste is the largest component of waste, accounting for 20-50% of total waste generated by countries. Due to the high cost of disposal, recycling construction waste is not an option but a necessity.
Innovation and Uniqueness
Why Our Project Stands Out:
Image generation AI has demonstrated powerful capabilities, but it requires high computing power. To overcome this, we applied LoRA (Low-Rank Adaptation) to an LCM model architecture, which is capable of few-step generation. By providing information about drawings and buildings during the fine-tuning process using LoRA, we showed that we can output design drawings based on information about the desired building.
Insights and Development
Learning Journey:
The development of this idea required extensive research. To understand the recycling process of construction waste in Korean society, we investigated information on construction waste treatment companies, unit prices of construction materials and recycled materials, and laws related to architectural design. Based on this research, we were able to develop the idea of using generative AI to further promote the recycling of construction waste.
Development Process:
The models used in our project (LCM-lora, BLIP) utilized Huggingface's diffuser and transformers libraries. Since there was no text data in the used dataset and training was conducted with limited computing resources, we built a computationally efficient pipeline using LoRA for the image captioning model and LCM model. The trained model can perform inference by only loading the LoRA module weights into Stable-Diffusion v1.5.
