Neural networks/machine learning

I think neural networks are going to play an integral part in terrain generation in the future.

Here is a neat paper displaying one of the best uses of machine learning in terrain generation at the moment https://dl.acm.org/doi/pdf/10.1145/3355089.3356553

I think that using real world terrain data to train neural networks could develop some incredibly helpful tools, such as supersampling


or content fill

Now I imagine that for Stephen getting into this wholly new area of neural networks, with all the features already on the roadmap for the next WM releases would be too much of an undertaking, but maybe you (stephen) could consider getting some help from someone well versed with machine learning, who would work separately on those features.

This is just me rambling, I know this is very unlikely to happen, but I can also imagine that being the industry first terrain software with neural network features would make some pretty big waves in the industry, as well as being an excellent marketing pitch. Especially since not even the big texturing software giants like Substance and Quixel have dipped their toes in the neural network waters (except for Substance Alchemist, which has the very limited use case of turning photos into PBR surfaces).

1 Like

Could also become a community project, letting users upload their world data to train the algorithm with.

Except for supersampling and feature adding, are you aware of other applications by chance?

Another feature I can think of right now would be texture synthesis, you could for example choose an area of the world wich suits your terrain look and it would synthesize a texture that fits the biome you chose.

The biggest issue here would probably be finding a source of royalty free high resolution satellite imagery to use as training.

1 Like