search

sherwinbahmani / 4dfy

4D-fy: Text-to-4D Generation Using Hybrid Score Distillation Sampling
https://sherwinbahmani.github.io/4dfy/
Apache License 2.0
288 stars 7 forks source link

4D result issue #6

Closed yuelei0428 closed 6 months ago

yuelei0428 commented 6 months ago

The unicorn in the 4d result does not move much. Is it a usual thing that occurs in the training process? What may I do to fix the problem?

The result link is here: https://github.com/sherwinbahmani/4dfy/assets/93293009/b9b1cf96-7fab-461d-ae34-77fd1203f4ee

sherwinbahmani commented 6 months ago

Hi,

Which config did you use?

I noticed that the low vram config with the threestudio extension causes this, while the full version does not cause this. From my experiments it is mainly down due to the stable nerf volume renderer in the config. I have not figured out why it is happening though. So, currently, to reproduce the results, a 80 GB machine may be required for the last stage. Or at least 48 GB.

sherwinbahmani commented 6 months ago

I updated the code with low memory config files, please use them as you probably trained this with the low memory config of the threestudio extension, which had some issues with motion