What's the purpose of custom mipmap build function.

[saedrna]

I have noticed that the code use a custom mipmap build function texture2d_mip for the texture sampling. I am wondering what's the difference with the built-in function in diffract, e.g., texture_construct_mip.

https://github.com/NVlabs/nvdiffrec/blob/e7f2181b8a60eb8fedcdb4ad4d05bff3c0cf9bc1/render/texture.py#L21-L32

I have tried to replace the lines below with texture_construct_mip, but it seems that there is no difference.

https://github.com/NVlabs/nvdiffrec/blob/e7f2181b8a60eb8fedcdb4ad4d05bff3c0cf9bc1/render/texture.py#L65C5-L76

[JHnvidia]

Hi @saedrna,

It's mostly equivalent, and the main reason is legacy. In previous work (https://github.com/NVlabs/nvdiffmodeling) we did experiments on shader filtering by computing custom mip-pyramids. For example, it's usually a good idea to artificially reduce material roughness to supress highlight-flickering at coarse miplevels.It makes more sense for level-of-detail, though, so we're not using it in the nvdiffrec codebases.

The texture_construct_mip function creates an opaque object with the mipmap information, so it doesn't support passing the individual miplevels to a torch optimizer.

[saedrna]

For example, it's usually a good idea to artificially reduce material roughness to supress highlight-flickering at coarse miplevels.

The texture_construct_mip function creates an opaque object with the mipmap information, so it doesn't support passing the individual miplevels to a torch optimizer.

Thanks, I get the idea. Probably similar to the purpose of prefiltered specular envmap, i.e., more rough object will be sampled from coarse miplevels as below.

            # Roughness adjusted specular env lookup
            miplevel = self.get_mip(roughness)
            spec = dr.texture(self.specular[0][None, ...], reflvec.contiguous(), mip=list(m[None, ...] for m in self.specular[1:]), mip_level_bias=miplevel[..., 0], filter_mode='linear-mipmap-linear', boundary_mode='cube')