Training GaussianPro using a GPU with 12GB VRAM

kcheng1021 / GaussianPro

[ICML2024] Official code for GaussianPro: 3D Gaussian Splatting with Progressive Propagation

https://kcheng1021.github.io/gaussianpro.github.io/

MIT License

651 stars 44 forks source link

Training GaussianPro using a GPU with 12GB VRAM #67

Open caioviturinofs opened 5 days ago

caioviturinofs commented 5 days ago

Is there anyway to train this GaussianPro with a GPU with 12GB of VRAM?

Thank you in advance :)

kcheng1021 commented 2 days ago

Hi, many thanks for your contribution first. For your question, first make sure 3DGS originally can train your scene with 12GB. Then you can downscale the training images size, or increasing the size of --percent_dense parameter (to make gaussian densify less), or decrease the propagated Gaussians' number (in line 440-442 of gaussian_model.py). Propagation module itself does not cost too much memory.