Language Modality Model for 3D Spacial Reasoning

Our research report first discusses the Language Modal Model (LMM) and its working principle through data collection and relevant research investigation, covering the definition, working process, development history, cutting-edge research, advantages, challenges and goals of LMM.

After gathering the basic information, we designs a series of generative AI dialogue experiments, trying to get a specific insight into the AI understanding of architectural space using the input of natural language and architectural picture space. We evaluate the responses of GPT4 from the perspective of architecture education and vocation as well as the reasonableness of its suggestions and design improvements. By conducting conversational experiments with generative AI, we explores AI’s understanding capabilities in the field of architecture design and the theoretical possibilities for integrating LMM in the "Architect-Free" Workflows. This aspect highlights the potential and limitations of AI in architectural spatial understanding.

Then, we uses a pre-trained model to conduct unit stacking design to study the autonomous capabilities of AI in building form design. It emphasizes the autonomy and assistance of AI in architectural design, as well as the role of AI in this field. By pre-designing the pre-trained model and converting the language input, Chat to Building used the model to conduct a simple unit stacking design, which reflects that the AI independent design has better control over the general shape of the building and has a more appropriate spatial understanding, but still Human participation is required for precise control, and AI-assisted design is very successful. 

Although our attempts are simple and insignificant, we can see the possibilities that different city and space models were produced by extracting design elements and adding a large number of vector design requirements. We hope that all assets generated by AI can be parameterized and controlled instantly and completely integrated with traditional workflows, and AI-assisted design can be generalized in the application for universal architectural construction practitioners - with lower threshold using the better understanding natural languages.

Keywords: Language Modality Models, Generative AI dialogue experiments, "Architect-Free" Workflows, natural language processing, parametric model production