The encoder in the paper is different from your implementation!

[qinglew]

You didn't use the graph-based encoder the paper described.

[yhbian]

Grateful for this neat pytorch implementation for FoldingNet!

However, I encounter the same issue. I train 300 epochs with several losses(like chamfer, earth-mover, or weighted version of both). None of them exhibits a satisfactory result. I guess the PointNet-like encoder might fail to capture the local geometry of the point cloud, which may provide crucial clues for the succeeding folding process.