Point Cloud Optimization

[v4r4rth]

Hi - I decided to move the issues to a new thread as the last one was mostly just me figuring out how to play with your nodes! I did manage to improve the quality of the splat by refining some parameters but the point cloud itself is not optimised for marching tetrahedrons - below I show the splat, along with the point cloud and generated mesh (without texture) -

My question is, can we find a way to remove the points which are not contributing to the RGB values? Specifically, can we eliminate the Alpha value points from the cloud?

I also came across and implemented this project, SuGaR (check this issue) which also suggests a method to extract textures and meshes from Gaussian Splatting which is very efficient (tested on provided and my own datasets) - I do hope you take a look once because it did work very well at extracting meshes from gaussian splats :) - The UV Maps were bad, but still, the quality was decent.

[watsieboi]

I ran into the same issue - my janky approach to solving was to add a remove background node (https://github.com/Jcd1230/rembg-comfyui-node) after the Zero123 image generation step. Then in the gaussian splatting workflow the image masks will actually be populated and the black background will not contribute to the splats.

[MrForExample]

Thanks for the effort friend, I am aware of SuGaR, and will definitely take a closer look at their method. I also adopted a better 3DGS to Mesh Conversion Algorithm in yesterday's update, it use nerf to fit 3DGS and then using marching cube to produce mesh & texture by fitting the nerf, and it works much better & faster, feel free to give a try