search

ActiveVisionLab / gaussctrl

[ECCV 2024] GaussCtrl: Multi-View Consistent Text-Driven 3D Gaussian Splatting Editing
https://gaussctrl.active.vision/
BSD 3-Clause "New" or "Revised" License
80 stars 6 forks source link

The training cost #1

Closed leyangjin closed 3 months ago

leyangjin commented 3 months ago

Hi, nice work! Just curious about the type and number of GPU used for training, and the time needed for the original training phase.

jingwu2121 commented 3 months ago

Hi, the entire editing process is an inference process, it doesn't involve any training, the only training here is the 3DGS training. As we said in the README file, we use an NVIDIA RTX A5000 GPU (24G) for training (editing and train 3DGS). To fit a smaller GPU, you can also decrease the batch size and the number of reference frames, it should also work. Our batch size is 3, with 4 reference frames (~22G). So the entire training time is around 9mins as we introduced in our paper if we edit 40 images, (which is a number that you can set in the hyperparameter, as introduced in the README file.) You can also decrease this number to speed up, should also work fine

leyangjin commented 3 months ago

Hi, thank you so much for your help!