[Paper Review] Generating Daylight-driven Architectural Design via Diffusion Models

Pengzhi Li, Baijuan Li|arXiv (Cornell University)|Apr 20, 2024

Impact of Light on Environment and Health4 citations

TL;DR

This paper proposes a novel AI-aided architectural design framework that generates daylight-driven architectural designs from massing models using diffusion models. It integrates parametric massing generation, a learned daylighting strategy via LoRA-fine-tuned diffusion models on a custom daylighting map dataset, and conditional text-to-image generation with ControlNet and GPT-4 prompts, achieving efficient, aesthetically coherent, and functionally informed design proposals with strong user validation.

ABSTRACT

In recent years, the rapid development of large-scale models has made new possibilities for interdisciplinary fields such as architecture. In this paper, we present a novel daylight-driven AI-aided architectural design method. Firstly, we formulate a method for generating massing models, producing architectural massing models using random parameters quickly. Subsequently, we integrate a daylight-driven facade design strategy, accurately determining window layouts and applying them to the massing models. Finally, we seamlessly combine a large-scale language model with a text-to-image model, enhancing the efficiency of generating visual architectural design renderings. Experimental results demonstrate that our approach supports architects' creative inspirations and pioneers novel avenues for architectural design development. Project page: https://zrealli.github.io/DDADesign/.

Motivation & Objective

• To address the lack of daylight integration in AI-generated architectural design by introducing a daylight-driven strategy.
• To streamline the architectural design process by enabling end-to-end generation from massing models to visual renderings.
• To enhance design efficiency and creativity by combining large-scale language models with text-to-image diffusion models.
• To create a novel dataset of computational daylighting maps for training architectural diffusion models.
• To improve alignment between generated architectural designs and the original massing geometry using conditional controls.

Proposed method

• Massing models are generated via a parametric algorithm in Grasshopper, using iterative volume addition and subtraction with random parameters.
• A custom dataset of daylighting maps is constructed using Honeybee plugin in Grasshopper, based on sectional profiles of massing models.
• LoRA-adapted Stable Diffusion v1.5 is fine-tuned on the daylighting map dataset to generate accurate window layouts.
• GPT-4 is used to generate diverse architectural text prompts categorized into types, styles, landscapes, and materials.
• ControlNet is integrated into the text-to-image diffusion pipeline to condition generation on massing geometry, ensuring shape fidelity.
• The final architectural renderings are produced by feeding GPT-4-generated prompts into the conditional diffusion model with geometry control.

Experimental results

Research questions

• RQ1 Can diffusion models be effectively guided to generate architecturally coherent façades based on daylighting performance?
• RQ2 How can a custom daylighting map dataset be constructed to support AI-based architectural design?
• RQ3 To what extent does combining GPT-4 with text-to-image diffusion improve design diversity and relevance?
• RQ4 How well do architects perceive the efficiency and usability of an AI system that integrates daylighting and form generation?
• RQ5 Can conditional control via ControlNet maintain geometric consistency between massing models and generated architectural renderings?

Key findings

• User studies show 88% of architects rated the method as significantly improving design efficiency.
• 84% of architects reported increased design inspiration when using the proposed framework.
• 72% of users found the architectural designs to be rational and well-proportioned, indicating strong functional coherence.
• The daylighting-driven façade generation strategy was positively evaluated by 64% of users, with only 14% finding it neutral or ineffective.
• The framework successfully generated diverse, visually coherent architectural designs across multiple styles and typologies, with filtered outputs ensuring plausibility.
• The integration of GPT-4 prompts with ControlNet-enabled diffusion models produced high-fidelity renderings closely aligned with the input massing geometry.