Question about generated 3DGS/2DGS scale.

[RUiN-jiarun]

Hi @bluestyle97 , thanks for your great work!
My question is, since the network is scale-agnostic, it predicts the gaussians straight-forward through the Transformer module. Is it possible to scaling the generated 3DGS/2DGS directly or through the network?
For example, I want to apply a scaling factor to the generated 2DGS object, now I can only apply the scaling factor to each gaussian attributes. The result is always obviously degraded. Any suggestions?

[bluestyle97]

Hi, I think you need to scale the position and scale attributes of the generated Gaussians simulteneously.

[RUiN-jiarun]

Oh yes I did apply scaling factor to both position and scale. I used the inverse sigmoid function and min-max scaling for scale attribute according to your sigmoid activation. e.g. if I scale the 2DGS for 3x, there will be some holes between gaussian primitives, thus made the appearance of the rendering results bad.
Any other methods to improve the results? Or is there any method to generate a scaled GS object from the network?