I had an amazing idea for tree and vegetation rendering. Just sharing here because i dont have time to experiment.
I read an article back in the summer about Unreals new voxel based tree renders. The leaves and branches are instanced and also voxelized. The voxel representation is used to simplify rendering at a distance, i.e. when the leaves fit inside a pixel.
I couldnt help thinking that its a shame voxels dont line up perfectly to pixels from all angles, for rendering in a screen space quad tree.
Then i thought maybe they could be a type of bounding heirarchical volume - sphere trees. The spheres always map the same way to pixels … theyre off, but at a distance the sphere beomes a point. The number of leaves in each sphere determines the point color.
Then for better sphere coloring, perhaps it is well solved with multiple regression or a neural net. The color data is first grabbed by rendering the brute force tree in different distances and with different lighting, and any other variables also contribute to make a data point. Then a sweep over the settings is done, depending on the number of variables this could take several minutes with spacial data points (i.e. location, rotation) gained by repeating the object. The multi dimensional parameter space is then fitted with a regression surface, or a neural net, then the result is hopefully a simple curvi-linear function that returns the best blob/disc color given the leaves inside. Perhaps this could be performed for a variety of preconfigured leafy variations for each tree type, as there may be no one function fits all solution.
I think maybe this could work the same way as voxels, but if each voxel is already rendered as a sphere, then there is no advantage and it is down to almost tbe same thing. Perhaps bounding sphere trees look nicer and are easier to construct though.
I understand some people feel strongly about AI, but this usage of a neural net would be a multi-variable data fitting application. It is not AI, it is a mathematical thing, and it is how many difficult or weird mathematical puzzles got solved in physics for the last couple of 100 years or so, minus the additional regression dimension that the neural net fits.