search

jenkspt / multires-hash-encoding-pytorch

Pytorch implementation of data structures and modules from InstantNGP
MIT License
26 stars 3 forks source link

Numerical Comparison bewteen the implementation and Instant-NGP #1

Open wen-yuan-zhang opened 1 year ago

wen-yuan-zhang commented 1 year ago

Thanks for your work! I am well aware of the difficulties involved. Have you tested the results of this implementation? I'm wondering whether this implementation is able to achieve the same numerical results as the raw implementation of instant-ngp?

jenkspt commented 1 year ago

Hi @zParquet - I haven't compared this implementation directly to https://github.com/NVlabs/instant-ngp. The only experiments I've done are with NeuS (https://github.com/jenkspt/NeuS/blob/hash/models/hash_fields.py). If you're looking for a pedagogical implementation of the multires hash encoding - then look no further. However a lot of other works have come out since I working on this. If you're looking for something complete, but more hackable than instant-ngp, I would recommend nerfstudio: https://github.com/nerfstudio-project/nerfstudio, which has a full reimplementation of instant-ngp nerf training. Also check out https://github.com/JamesPerlman/TurboNeRF-Blender if you're experienced with C++ and CUDA. Both of these are permissively licensed unlike instant-ngp. Hope this helps!

wen-yuan-zhang commented 1 year ago

Thanks for your reply! Following your advice, I find that nerfstudio reimplemented instant-ngp using pytorch which may be helpful. I'll keep an eye on this problem. Thanks again!