Dec 31, 2023 · When you have a training image which captures a far-away gaussian, the world-space 3D mean's gradient of the far-away gaussian should be small because perturbing one unit of the GS …

Nov 18, 2023 · I face the same problem, by checking the code, I think the main reason leading to this problem is the "pruning". Sometimes, all Gaussians are pruned so that the number of gaussian is 0, …

Dec 29, 2023 · Adaptive Gaussian control uses partial derivative (dL/du) to determine which Gaussian to be split/cloned. The function precomputes the gradient to every Gaussian in the visible frustum.

Mar 20, 2025 · Notifications You must be signed in to change notification settings Fork 2.6k

Oct 25, 2024 · Dear Author, Why is it said that the view-space positional gradients are calculated in the NDC coordinate system? I noticed in the backward.cu code that the distance d is calculated in the …

Nov 21, 2023 · I am recently looking at the backward pass of the gradient calculation. It seems to me that the gradient passed to the transmittance $T$ term is missing, which affects the calculation of the …

May 17, 2025 · In diff-gaussian-rasterization/cuda_rasterizer/backward.cu::computeCov2DCUDA(), the gradient of the conic matrix (dL_dconic) is incorrectly initialized by accessing float3 dL_dconic = { …

Hello, I checked the code and found that there is an all-zero matrix screen_point (means2D), which is used to store the gradient of the two-dimensional Gaussian ellipsoid, but when propagating forward, I …

Nov 2, 2023 · Hello, I am trying to improve the gaussian rendering code. However, when I use the torch.autograd.gradcheck to verify the correctness of the gradient computed by the cuda code, all the …

Dec 5, 2024 · the gradient computation should be correct. However, in my humble opinion, it will be more intuitive to treat conic2D as three numbers instead of a float4 with one item empty and double …