search

cvachha / instruct-gs2gs

Gaussian splatting implementation of Instruct-NeRF2NeRF: Editing 3D Scenes with Instructions
https://instruct-gs2gs.github.io/
MIT License
62 stars 1 forks source link

very slow training speed using scannet++ dataset #15

Open ZiyangYan opened 2 months ago

ZiyangYan commented 2 months ago

Hi, I apply igs2gs using a scannet++ dataset but the training speed is very slow (GPU: 40G A40). The checkpoint I used is from splatfacto with 30K iteration. image

cvachha commented 1 month ago

Hi, I'm thinking this is running slow due to Instruct-Pix2Pix. Instruct-GS2GS applies Instruct-Pix2Pix on every image in the training set (for the first [num of images] iterations) which ideally should have an image resolution of 512x512 or less (larger images run slower). Instruct-GS2GS will train the splat for 2.5k iterations after editing each input image.

If the image resolution is large, you can downscale it by adding nerfstudio-data --downscale-factor {2,4,6,8} to the end of your ns-train commands.

Let me know if this helps with the speed.