Harvard University, Graduate School of Design
Neural Bodies, Spring 2022
Instructor: Andrew Witt
Neural Bodies, Spring 2022
Instructor: Andrew Witt
What would a crowd-sourced Chinatown gate entail?
The project aimed to utilize Chinatown gates located around the world as a dynamic heritage. A custom dataset was created by scraping images from Instagram using the hashtag #chinatowngates. This dataset was then used to train a pix2pix model, which was compared with a "facade" dataset commonly used in architectural image training. The project also included experimentation with different generative models such as StyleGAN and MiDaS (Depth mapping) to create images of Chinatown gates. A paired dataset was prepared specifically for the Chinatown gates and a model was trained on it. Finally, an attempt was made to translate the resulting images into 3D gates.
Image Scraping
Testing
StyleGAN - Facades dataset
MiDaS - Depth Mapping
3D visualization of generated images
pix2pix Paired Dataset
Results
