Hacker News Reader

mdp2021

> an art form

Which is why the linked Daniel Huffman's technique employed the convincing results of Blender.

The biggest thing that I'm seeing that Singh has over Huffman is the adjustable detail-preserving terrain smoothing. Detail-preserving blur algorithms don't generalize well. Application-specific NN solutions tend to excel at this problem, by having a model of what details are important vs incidental.

I would like to see a combination of the two, with the terrain pre-processed before being used in Blender.

This is a similar problem as automatic LOD computation for 3D game assets though, and that has been done for a long time without ML (but maybe ML may actually help there too, no idea).

I still find the 'human touch' argument a bit unconvincing for this type of problem.

ML is absolutely being used in automatic LoD computation for far better results, as well as in other geometry transformations (with Unreal Engine introducing the generic "ML Deformer", for example).

> Application-specific NN solutions tend to excel at this problem, by having a model of what details are important vs incidental

I'd say this is a context in which "the problem of Transparency" becomes evident: we would like to "see why", to come to learn a perfected deterministic algorithm from the implicit model that the NN developed.

Reply on news.ycombinator.com