Detial of regularizing 2dgs

Stable-X / StableNormal

[SIGGRAPH Asia 2024 (Journal Track)] StableNormal: Reducing Diffusion Variance for Stable and Sharp Normal

Apache License 2.0

518 stars 23 forks source link

Detial of regularizing 2dgs #15

Closed Binyr closed 2 months ago

Binyr commented 2 months ago

Hi,

I use your model to inference normal maps of DTU dataset. And regularize 2dgs normal map after iter 7k with lambda 0.05. The loss is similar to the 2dgs depth_normal regularizing loss. Finnally, I get a slightly different results shown below. Could you provide some details to help me reproduce these results?

Thank you! Yanrui

hugoycj commented 2 months ago

Hi Yanrui, thanks for providing your reimplmentation result. Over the past three months, we've iteratively updated StableNormal from v0.1 to v0.3, incorporating a more diverse dataset and modifying our pretraining and training techniques. As a result, the 2DGS outcomes may differ significantly from previous versions.

Regarding regularizing loss, you can refer to this PR for details. Our regularization method differs slightly from your implementation, which could also account for some performance variations.

seulqxq commented 1 month ago

Hi, can I ask you how to combine stablenormalX with 2dgs?

hugoycj commented 3 weeks ago

@seulqxq We have uploaded the train with normal scripts to https://github.com/hugoycj/2d-gaussian-splatting-great-again/blob/main/scripts/train_wmask_wnormal.sh. Feel free to have a try

seulqxq commented 3 weeks ago

@hugoycj Thank you very much. I will undertake an attempt.