Closed Lucklycat closed 1 year ago
Hello,
Deforming a sphere to a character (or elephant) is quite challenging. For those examples in the paper, we used a sphere with 12k triangles (higher tessellation than the sphere provided in the code repo) as an initial guess for the rasterizer, and optimized at a resolution of 2048 × 2048 pixels for 10k steps. You can tweak the Laplacian regularizer weight to steer the optimization process a bit (e.g., to better capture smaller features).
That said, you would likely obtain better results with a better initial guess that a sphere, or by switching to nvdiffrec, which handles this case without the need for an initial mesh. https://github.com/NVlabs/nvdiffrec
Thanks for your reply, I'll try
I read your paper and video in detail,about LEARNING MESH AND MATERIALS FROM IMPLICIT SURFACES,How to improve the effect? Can you give me some suggestions on setting super parameters? I used my own model as the target, started from the sphere you provided, and carried out experiments. I found that the feet and face were always unsatisfactory? There is a big gap with the rendering of snails and elephants mentioned in your paper. can you give me some adivce.![image](https://user-images.githubusercontent.com/17697877/209915438-7aa87576-cef4-41f8-b1a0-df8c372beff9.png)