Nikhil Mohan

Hi @evonneng could you try the following solution https://github.com/nerfstudio-project/nerfstudio/pull/910 and let me know if this fixes your issue?

I have outlined a proposal in this PR - https://github.com/nerfstudio-project/nerfstudio/pull/886 . This is without the is_thing and loss_weight above, but can be added in pretty trivially if required

So Ive been debugging this issue on my side - and have been able to reproduce / remove the following error through the changes here - Though im not really...

> Tested to 30k iterations with eps=1e-3 and it seems to work Yeah this seems like a better solution - it works for me as well, Ive reverted the jit...

> Would be good to do a quality comparison to make sure it doesn't affect results. Here is what I get when testing on the poster dataset - orange is...

> This looks fine to me. I think we should still suggest that people pre-scale their images. We actually used to have this feature and found that people would upload...

> Awesome thank you for doing this! With the depth rasterization, I think it would make sense to pack the RGBD into a float4 array, rather than a separate float*...

> I think it might make sense to have a separate function for rasterize 4-dim. We have a rasterize 3-dim (for rgb) and N-dim, but we could also have one...

> Hi @nikmo33, I don't think there are any plans right now, but if you would like to work on this together I would be happy to collaborate. I think...

I also see that there is work to support alpha supervision which is great! https://github.com/nerfstudio-project/gsplat/pull/70 . I think something very similar could work in this case too?