search

ashawkey / RAD-NeRF

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

Question on the sampling steps used in RAD-NeRF #22

Closed flyingshan closed 1 year ago

flyingshan commented 1 year ago

I notice that the param "max_steps" is set to 16 in RAD-NeRF, which makes sampling more rays in a batch in training feasible, while in original torch-ngp, it is way larger. My question is that if "max_steps" is naturally smaller in the face rendering scene, or other improvements are made in the code. Looking forward to your reply!

ashawkey commented 1 year ago

@flyingshan Hi, human head is much simpler then natural scenes, so it needs naturally fewer steps to render.

flyingshan commented 1 year ago

Got it, thank you for your reply!