CityGPT: Generative Transformer for City Layout of Arbitrary Building Shape

Rongjing Xie, Wenyu Han, Shuhang Ge, Congcong Wen, Chen Feng

Primary Area: generative models

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: City layout generation, GPT, Infinity generation, Polygon layout

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We provide a transformer-based model and a two-stage masked training pipeline for city layout generation tasks of arbitrary building shapes.

Abstract:

City layout generation has gained substantial attention in the research community with applications in urban planning and gaming.
We introduce CityGPT, the generative pre-trained transformers for modeling city layout distributions from large-scale layout datasets without requiring priors like satellite images, road networks, or layout graphs. Inspired by masked autoencoders (MAE), our key idea is to decompose this model into two conditional ones: first a distribution of buildings' center positions conditioned on unmasked layouts, and then a distribution of masked layouts conditioned on their sampled center positions and unmasked layouts. These two conditional models are learned sequentially as two transformer-based masked autoencoders. Moreover, by adding an autoregressive polygon model after the second autoencoder, CityGPT can generate city layouts with arbitrary building footprint shapes instead of boxes or predefined shape sets.
CityGPT exhibits strong performance gains over baseline methods and supports a diverse range of generation tasks, including 2.5D city generation, city completion, infinite city generation, and conditional layout generation.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 6752