Károly Zsolnai-Fehér contrasts fast noise-based terrain generation with learned methods that can capture more organic terrain distributions but become expensive when every region depends on the whole world. The research he reviews combines learning with local generation and shows examples in Minecraft as well as continuous terrain.
Károly Zsolnai-Fehér explains how overlapping windows denoise a requested region and blend their predictions. Limiting the computation to nearby windows keeps a query's cost from growing with the total generated world. A multiscale Laplacian technique separately handles large height differences and fine terrain features before combining them.
Károly Zsolnai-Fehér reports that the research trained in two weeks and ran interactively on an older consumer GPU, with code and a Minecraft mod available. Those are results presented from the research, not an independent reproduction in this video. The useful takeaway is the combination of local consistency and multiscale detail; the closing Lambda sponsor read is omitted.
Watch on YouTube
