Terrain Diffusion Network: Climatic-Aware Terrain Generation with Geological Sketch Guidance

Authors: Zexin Hu, Kun Hu, Clinton Mo, Lei Pan, Zhiyong Wang

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive experiments on a new dataset constructed from NASA Topology Images clearly demonstrate the effectiveness of our proposed method, achieving the state-of-the-art performance.
Researcher Affiliation Academia 1The University of Sydney 2Civil Aviation Flight University of China
Pseudocode No The paper describes its methodology in narrative text and through a diagram (Fig. 2) but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/TDNResearch/TDN.
Open Datasets Yes A new dataset is constructed from NASA Topology Images (Allen 2005) to evaluate the effectiveness of the proposed TDN.
Dataset Splits Yes We obtained 10,446 samples for training and 2,611 for testing.
Hardware Specification Yes We used one NVIDIA RTX 3090 GPU card to train our model.
Software Dependencies No The paper mentions 'Pysheds (Bartos 2020) packages' but does not provide specific version numbers for this or any other software libraries or dependencies required to replicate the experiments.
Experiment Setup Yes The U-Net like synthesisers in TDN comprise 3 downsampling blocks, 3 upsampling blocks, and a middle block with 8-head selfattention mechanisms employed for embedding the diffusion step t. With a total of 1.216 billion parameters, the model s training is conducted with a learning rate of 1.0e05 and a batch size of 6. During inference, TDN takes a set of user sketches as the input and iteratively generates a noise-free terrain latent representation. In total, 36 steps are taken to derive the final latent representation.