Matthew Tancik

Can you share some example data, it may help with debugging.

Oh, I'm only now realizing that the sf dataset is one what we provided. This dataset is bad and has been removed in the latest release.

Also consider raising an issue on the TinyCudaNN github. They may have more pointers - https://github.com/NVlabs/tiny-cuda-nn/issues

Can you provide example example renders. Also a screenshot of the camera distribution (from the visualization in the viewer).

If you are able to recapture the scene, you should try rotating as you walk through the scene in increase the number of angles that you view the scene.

If it works works with NGP it should also do reasonably well with nerfacto. Can you share an example of your results to make debugging easier?

I'm able to replicate this error. We would also see this issue when we trained nerfacto for a large number of iterations (>50k). Seems like it appears early for this...

@snarb We are using TinyCudaNN for our network and encoding, this is implemented in half-precision. The `mixed_precision=True` enables gradient scaling (https://pytorch.org/docs/stable/amp.html#gradient-scaling) which we find necessary when using this library. Are...

I think is fixed in the latest version. If you don't want to update, try making the change in this PR https://github.com/nerfstudio-project/nerfstudio/pull/910

Some variance is expected to some extent due to the random nature of the optimization. "Sigificant" differences is a bit surprising and should be looked into a bit more. I...