search

MrTornado24 / Next3D

[CVPR 2023 Highlight] Next3D: Generative Neural Texture Rasterization for 3D-Aware Head Avatars
https://mrtornado24.github.io/Next3D/
470 stars 29 forks source link

Issue of reload_modules #15

Closed zxYin closed 1 year ago

zxYin commented 1 year ago

Hi,

I have been trying to use the generator TriPlaneGenerator. When loading it with legacy, everything is fine. However, I am experiencing some issues when reloading the network pickle into it by using copy_params_and_buffers , as shown in the demo code. I have noticed some artifacts when doing so, which are visible in the following image:

1

I also came across what seems to be a bug in the code at this location: https://github.com/MrTornado24/Next3D/blob/bdab84fc12cb61bd1002a3efb2c0a707a1139d97/training_avatar_texture/triplane_next3d.py#L226

I would appreciate any insights or assistance you could provide to help me resolve this issue.

Thank you for your time and support. I look forward to your response.

Piurvus commented 1 year ago

Hey, I had the same issues. It seems as if their code is somewhat faulty. However, thanks to Nvidias persistence class you can get their previous code. G['G_ema']._orig_module_src

RaymondJiangkw commented 1 year ago

Hi, you can see this. 👀

zxYin commented 1 year ago

Thanks for your help