search

ayaanzhaque / instruct-nerf2nerf

Instruct-NeRF2NeRF: Editing 3D Scenes with Instructions (ICCV 2023)
https://instruct-nerf2nerf.github.io/
MIT License
769 stars 64 forks source link

about iteration times #59

Closed trltzy closed 8 months ago

trltzy commented 10 months ago

My nerfacto model was trianed with default iterations 30k, and I set the --max-num-iterations of in2n as 3k, the training seems slow with a very long estimation time. So I have two questions: 1) can I just set a smaller interation time in instruct-nerf2nerf? 2) or I have to set the same interation time as I trained in nerfacto? which means if I want to have a quicker trainning I have to lower the interation of nerfacto first? image image

ayaanzhaque commented 10 months ago

You can adjust the total iteration count in the config, or with the flag --max-num-iterations 10000. However, if you are using the viewer, you can just have it open and pause/end training whenever you feel satisfied with the edit. The number of iterations to get a satisfactory edit can differ per scene/edit

trltzy commented 10 months ago

thx finally I ran up the project, but it seems the effects of my prompt in webviewer is hard to detect. My prompt is {"give it a cowboy hat"}, it looks very like my original model. Is there some parameters I should set additionally?

ayaanzhaque commented 10 months ago

Are there some edits at all?

ns-train in2n --data data/nerfstudio/face/ --load-dir {nerfacto model} --pipeline.prompt "give him a cowboy hat" --pipeline.guidance-scale 7.5 --pipeline.image-guidance-scale 1.3 --pipeline.model.near-plane 0.2 nerfstudio-data --scene-scale 2.0 --downscale-factor 2 Try with that command. Those are the parameters we used in our paper.

ayaanzhaque commented 8 months ago

Let's discuss this in #60