Is there anyway to resume training just for the texture?

NVlabs / nvdiffrec

Official code for the CVPR 2022 (oral) paper "Extracting Triangular 3D Models, Materials, and Lighting From Images".

Other

2.09k stars 222 forks source link

Is there anyway to resume training just for the texture? #78

Closed Holorei closed 1 year ago

Holorei commented 1 year ago

Hi and Thanks for your hardwork! After receiving the DMTET mesh optimization and first validation result, the process went out of memory in GPU for texture training, knowing that a parameter in the config file wasn't modified. I'm wondering if there is a way to resume training just for texture optimization.

jmunkberg commented 1 year ago

Thanks @Holorei ,

Yes, you can just run the second pass. For an example, please refer to the config https://github.com/NVlabs/nvdiffrec/blob/main/configs/spot_metal.json

Add the result from the first pass (available in the dmtet_mesh folder) as base_mesh and envmap, and start the optimization

Optionally, you can also use the "lock_pos" : true / "lock_light" : true options to disable geometry/light optimization if you are only interested in tweaking texture quality.