Training with custom dataset

[Mikey-E]

While training with a custom dataset, always at step 9100:

RuntimeError: Function _RasterizeGaussiansBackward returned an invalid gradient at index 2 - got [0, 0, 3] but expected shape compatible with [0, 16, 3]

Do you know if it's possible to fix this?

[Talnex]

You can check this issue
graphdeco-inria/gaussian-splatting#482

[Mikey-E]

In that issue I try following the advice and manage to get it to train using a radius of 32. The results are blurry. Left is ground-truth and right is 2dgs. Can you give any advice?