ashawkey / RAD-NeRF

Real-time Neural Radiance Talking Portrait Synthesis via Audio-spatial Decomposition
MIT License
878 stars 153 forks source link

What's new in Grid Encoder? #33

Closed yerfor closed 1 year ago

yerfor commented 1 year ago

Thanks for your brilliant work!

I'm interested in the modifications in the latest commit that update the grid encoder, which seems to replace the tilegrid to hashgrid, Can you briefly explain what's the motivation and how you achieve it?

Sincerely.

yerfor commented 1 year ago

I just ablated the modifications, and it seems that the additional .float() in ambient net contributes significantly to the HEAD_PSNR. By contrast, the grid_type or interpolation_type have no obvious improvements.

image
ashawkey commented 1 year ago

@yerfor Hi, it's just an updation to keep the code consistent with other projects, and also bug fixing for gridencoder. Thanks for the ablation!

yerfor commented 1 year ago

Thanks a lot. Your implementation of Encoders and RayMarching in CUDA is high-efficient and elegant. May I use your implementation in my project GeneFace, to replace the vanilla NeRF implementation?

ashawkey commented 1 year ago

Of course!

yerfor commented 1 year ago

thx!