Closed lcc815 closed 10 months ago
Hello @lcc815 ,
For the first line: It converged better with a gradient scaling of the light gradient. For the second line: We perform gradient scaling in the MLP code https://github.com/NVlabs/nvdiffrec/blob/main/render/mlptexture.py (that helped convergence), and the division is there to partly compensate for that scaling.
These hyperparameters were used to generate the results in the paper, but are heuristically chosen. Feel free to experiment with other values.
thanks for the soon reply :)
i am confused of these two lines of code: here and here why should the grad be rescaled?
Thanks,