I never took it beyond a simple dense network for a blade of grass, but I'd like to re-approach this using Tensorflow.JS. You can see the stupid simple POC at: https://jacob-ebey.github.io/gen-3d/ if you're interested.

Having experience working in CAD/CAM software, I have a feeling that the struggle in an approach like this will be maintaining meaningful output topology when adjusting parameters.

The simple approach in the grass POC was to have a single network input that was between 0 and 1 for each grass model. If we had 4 grass models, the training batch would have 4 entries with the first input being 0, the second being 0.25, etc... Then to generate a "grass" model, you could feed in any value between 0 and 1 to get a new model.

The catch here is that to maintain a meaningful output, the input models must have the same "shape". I.e the same number of vertices and indices as the indices from a base model are mapped to the new vertex values spat out by the network.