Question: The difference in gradients between interpolate and antialias

[yuedajiong]

1. After interpolation, good rendering results have been achieved, and we can directly use loss.backward() to obtain gradients. However, the gradients obtained are less accurate than backward after antialiasing. Why is this the case? For rendering speed, I would like to only perform rasterization and interpolation in the early stages of fitting. However, the gradients (including pos and col) are not very accurate. This means I must perform the antialias operation throughout the entire fitting process.

   "the real goal is to obtain gradients w.r.t. vertex positions related to occlusion, visibility, and coverage." ?

2. Can the gradients (including pos and col) obtained after antialiasing be used to guide the growth and pruning of vertices and faces?

3. To OpenDR, Rasterizer, PyTorch3D,..., and NvDiffRast, which gradients that generated by these algorihtms is 'better' for E2E(image-mesh-image') 3D fitting?

Thanks ！