Alpha weight formula

[martinwinter-huawei]

Is there a reason, why diff-gaussian-raserization uses

float weight = expf(-max_scale / depth ) + phi / (depth * depth) + phi*phi ;

while diff-gaussian-rasterization_ms uses

float weight = expf(max_scale / depth ) + phi / (depth * depth) + phi*phi ;

And in general, I think the paper does not go into detail here, why are these coefficients manipulated in such a manner to produce the weight, is there a specific reasoning behind it?

[martinwinter-huawei]

I just saw that this submodule has been removed in favor of diff-gaussian-rasterization_ms_nosorting, which uses the formula with the positive max_scale.
But the last question would still interest me 👍