The lr of camera optimizer will be ignored when loading from a checkpoint. Using large step in scheduler (max_steps=200000) leads to extra large lr on cameras, which leads to bad results when editing the scene.
Switching to a much smaller max_steps=5000 can ask it to always use lr_final
Hopefully this can address the issue https://github.com/ayaanzhaque/instruct-nerf2nerf/issues/60
The lr of camera optimizer will be ignored when loading from a checkpoint. Using large step in scheduler (
max_steps=200000
) leads to extra large lr on cameras, which leads to bad results when editing the scene.Switching to a much smaller
max_steps=5000
can ask it to always uselr_final