Autoregressive Generation of Static and Growing Trees

TL;DR

This paper introduces a multi-resolution transformer architecture for efficient static and growing tree generation, capable of processing complex structures faster and with less memory, and extends it to conditional and 4D tree generation tasks.

Contribution

A novel multi-resolution transformer architecture with skip connections that improves speed and memory efficiency for tree generation, including dynamic growth simulation.

Findings

Faster processing speed compared to vanilla transformers

Lower memory consumption during tree generation

Effective in generating complex and 4D trees

Abstract

We propose a transformer architecture and training strategy for tree generation. The architecture processes data at multiple resolutions and has an hourglass shape, with middle layers processing fewer tokens than outer layers. Similar to convolutional networks, we introduce longer range skip connections to completent this multi-resolution approach. The key advantage of this architecture is the faster processing speed and lower memory consumption. We are therefore able to process more complex trees than would be possible with a vanilla transformer architecture. Furthermore, we extend this approach to perform image-to-tree and point-cloud-to-tree conditional generation and to simulate the tree growth processes, generating 4D trees. Empirical results validate our approach in terms of speed, memory consumption, and generation quality.