Welcome to GAN for Urban Design project! It is a research on the use of Generative Adversarial Networks in the field of generative Urban Deisgn. Here, in particular, I have used a Pix2Pix model with the implementation from GANs Specialization.
Some of the results achieved during training with different models. The generated blocks are highighted with red color for the sake of clarity.
Arxiv | SimAUD Video Presentation (8 min)
For dataset generation refer to Urban Datasets repo
In order to create the datasets for training the model (or testing the existing model weights), please, refer to this repo. I have used the images with 256x256 dimensions.
You can test the model or start your training from the weights of the already trained models:
Please, configure the input parameters (save directory) in the config.py.
$pip install -r requirements.txt
$python generate.py images model.pth
- style corresponds to the style of training images (see the illustrations above if using on eof pretrained models)
- white and empty central block ("construction site")
- surroundings present in a large part of the image
- scale 1:3000
- image dimensions 256x256
Coming soon
Bibtex format:
@inproceedings{gan4ud,
author = {Fedorova, Stanislava},
title = {GANs for Urban Design},
year = {2021},
month = {04},
pages = {9},
booktitle = {In proceedings of 12th Symposium on Simulation for Architecture and Urban Design (SimAUD 2021)}
}