Question about the final result

[Wakals]

Thanks for your work for 3D generation, and I have learned a lot from your project. However, after I followed the README, and optimized 'a fox', ''' python launch.py --config configs/gaussiandreamer-sd.yaml --train --gpu 0 system.prompt_processor.prompt="a fox" system.sh_degree=3 ''' I got a collapse result. There's noting rendered in ./output/a_fox@.../save/it1200-test.mp4 (the video is pure black) and I cannot open the ./output/a_fox@.../save/it1200-test-color.ply by MeshLab (It warns me that mesh contains 63539 vertices with NAN coords and 0 degenerated faces Corrected.). So may there anything wrong for my operations? I'll appreciate it if you reply!

[pramishp]

Getting the same issue ! However, the shape.gif and shape.ply are fine.

https://github.com/hustvl/GaussianDreamer/assets/15194546/71704572-4cac-4b35-acd3-a3fc503a064c

it1000-train

shape.gif

[Zhao-Yihao]

I have met the same problem, the ply file cannot be import to unity engine.

[Peiqi00]

I also encountered the same problem, the rendered results .png and .mp4 are both black with no content. May I ask if you have found a solution?

[Zhao-Yihao]

Actually I didn't solve directly, but it may be because you did not install the pkg tiny-cuda-nn

[Wakals]

I also encountered the same problem, the rendered results .png and .mp4 are both black with no content. May I ask if you have found a solution?

Actually I have not tried to solve this problem. I find that the coordinates of GS do not update during the training (and sometimes go to NaN) and the loss keeps zero or NaN. What's I have learned from the author(s) is that this issue may be caused by gradient propagation for 3DGS within PyTorch-Lightning framework. And they are trying to solve it.

[Zhao-Yihao]

yes please check the if pkg tiny-cuda-nn installed