3D terrian generated using SinGAN.

• block2vec model comes from skip-gram and auto-encoder.
• The 3D space down/up sampling is based on grid sampling and trilinear interpoltation.
• re-implementation of SinGAN part with some fixes: 1. BatchNorm replaced by InstanceNorm, 2.reconstruct loss from MSE to Cosine Similarity.

3D terrian has a large number of available tokens: dirt, water, grass, tree, etc. A space with 100 x 100 x 20 blocks with 100 different tokens will cost 2e7 float bytes.
block2vec uses skip-gram and auto-encoder model to transform geo space into a latent token-based space with few dims.

fig. a list of tokens reprenseted in three-dimensional space

Down/Up sampling method using grid sample and trilinear interpoltation is happend in the latent space.

fig. down sample in latent space then reprensted in geo space

Random Generate

raw terrian

random noise generated 1

random noise generated 2

Random Size Generate
Can generate any size of terrian, due to full convolution network.

2x size