Thank you very much for your work.
(1)I also learned a lot from your recent work. However, the map in the mesh output by nvdifmodeling contains noise, which is still very troublesome.
As you told me in question https://github.com/NVlabs/nvdiffmodeling/issues/19, I understand that the noise in the texture here is because the high-resolution map is not used during training, so the gradient is not conducted.
Does that mean that if I keep the texture Res and train Is it possible to solve this problem if res is consistent? Or just reducing the noise can't be eliminated.
I tried to carry out experiments based on ewer, and set 512, 1024, 2048,respectively. I found that noise still exists.
At the same time, the higher the resolution, the whiter the map.So is the noise still related to the good training results? For example, can it be reduced by increasing the batch.
(2)Or these noises are not important, because according to the original paper, nvdifmodeling hopes to solve:
So as long as the optimization results are good, the final data will still be used in nvdifmodeling, and it is not recommended to use it in other commercial software, because some algorithms may not be universal?
Thank you very much for your work. (1)I also learned a lot from your recent work. However, the map in the mesh output by nvdifmodeling contains noise, which is still very troublesome. As you told me in question https://github.com/NVlabs/nvdiffmodeling/issues/19, I understand that the noise in the texture here is because the high-resolution map is not used during training, so the gradient is not conducted.
Does that mean that if I keep the texture Res and train Is it possible to solve this problem if res is consistent? Or just reducing the noise can't be eliminated.
I tried to carry out experiments based on ewer, and set 512, 1024, 2048,respectively. I found that noise still exists.![image](https://user-images.githubusercontent.com/17697877/217225655-01d2ae76-153a-4632-ae41-5993d0fe92b1.png)
At the same time, the higher the resolution, the whiter the map.So is the noise still related to the good training results? For example, can it be reduced by increasing the batch.
(2)Or these noises are not important, because according to the original paper, nvdifmodeling hopes to solve:
So as long as the optimization results are good, the final data will still be used in nvdifmodeling, and it is not recommended to use it in other commercial software, because some algorithms may not be universal?