Physical Model Enhancement Tool
-A Pix2Pix Approach-

Botao Sun & Zhian Yin

Final project for the MIT class 4.550/4.570 Computation Design Lab
Development: February-May, 2025
Instructor: Prof. Takehiko Nagakura & Dr. Daniel Tsai

Overview

Our project focused mainly on the early phase of design project, when study physical models are widely used. Current physical model based are mass-produced such as blue foams. For design references, architects use physical models to assist spatial description in a tangible form. Our approach explored a novel method to enhance physical model visualization to add real-time, multi-layer material and texture information from physical manipulation of the foam forms.

Methodology

We used Pix2Pix GAN as our generative model base, and take live video stream from camera as input to detect model parts in real life. Then we used line detection to isolate form contours, and finally applies line-to-image generation from the video feed to textured architectural massings for each frame. For this iteration we trainned our model only on brick surfaces, but we suggest that other building surfaces can be achieved through trainning the model on a more diverse dataset.

Fig 1. Conventionally Massing-to-Rendering visualization

Often limited to single-frame, fixed-camera angle visualization.


Fig 2. Proposed Real-Time Interactive Visualization

Animated, live-feed of building massing and texture.

Fig 3. Proposed Pipeline

Enhancing visualization using fast image generation from trained GAN model. Identify architectural elements in different colors and regions.

Closing-Thoughts

The real-time visualization we proposed aimed for efficiency to communicate design ideas been production and iteration of physical models and other visualization of building materials. Our approach can also be further applied to VR environment that allows interactive design among multiple parties. In addition, our structure can also be applied to support more technical aspects of building design. For example, a trained generative model on thermal or load analysis could also be incorporated into this translation between massing and structural properties.

Project Video

Project Demo Video

 


several
2022 All rights reserved.