Closed rowellz closed 1 week ago
What torch version are you using? There are problems with Torch 2.0.1. If so, you could upgrade to 2.1.2.
You absolute genius! I was indeed on Torch 2.0.1. After running pip install -U torch==2.1.2
it worked! Training did seem to be a bit faster, but tough to notice due to colmap still :/ That said, what was most impressive was that the splat took up half the disk space than the one produced from the previous nerfstudio version. Literally went from 5.4 MB to 2.7 MB which is so huge when trying to render these splats on the client. Congrats on the awesome work, I am looking forward to regenerating all of my splats now, lol
Closing this one for the related issue #3235
Splat quality significantly decreases after upgrading nerfstudio & gplat. I am using nerfstudio in a docker container based off of the nerfstuido/nerfstudio image after I ran into problems using the dromni/nerfstudio image a few months back.
Steps to reproduce the behavior: Upgrade nerfstudio & gplat with
pip install -U nerfstudio
Process images withns-process-data images
Process Splat withns-train splatfacto
Observe splat model in viewer. The statue is not recognizableExpected behavior The statue should be easily noticeable after loading the splat model
Statue splat before upgrading gsplat & nerfstudio
Statue splat after upgrading gsplat & nerfstudio
Both of these splats used the same source video, which can be viewed when clicking on the "Download Original Video" button in the splat viewer, which is at https://ez3d.nyc.
Thanks again to all the folks at nerfstudio for all of your awesome work. Please let me know if there is any more information I might need to provide to help resolve this issue.