search

google-research / nerf-from-image

Shape, Pose, and Appearance from a Single Image via Bootstrapped Radiance Field Inversion
Apache License 2.0
381 stars 18 forks source link

Can I enlarge the resolution from 32*32 to ...? #15

Closed 013292 closed 1 year ago

013292 commented 1 year ago

Thanks for sharing this work. May I ask the right or elegant way to enlarge the resolution. The default is 3232 right? Does it have any negative effect if I enlarge the resolution? Why do you choose the 32 32 in both this work and Pix2NeRF? Appreciate :-)

dariopavllo commented 1 year ago

Hi,

The default resolution is 128x128. I believe Pix2NeRF (who has been written by different authors) is trained at a lower resolution (64x64).

The choice of the resolution is mostly a matter of computational cost / available GPU memory, as well as stability during training. For some datasets (e.g. ShapeNet), 128x128 is the native resolution of the dataset. You might be able to increase the resolution, but most likely you also need to increase the strength of the R1 regularization to improve stability.