DiffRF: Rendering-guided 3D Radiance Field Diffusion (CVPR'2023)

[@dddd-wf6fn]

如果有很多万级带有语义的三维模型，是否可作为训练数据，用three.js加载后，自动输出训练数据？

[@dangthanhtuanit]

I wonder if there is an actual implementation since code is not available?

[@상수-u2i]

From my point of view, this paper seems to rely heavily on Plenoxel's voxel grid. It is considered an appropriate method for applying Diffusion to actual volume rendering. However, it is hard to find references to Plenoxel in paper. There is little information about Spherical harmonic in appendix. Can the method proposed in the paper be applied to other Nerf-like methods?

[@xandermasotto7541]

I had the same question. The way they fit the voxel representation seems to be the same as Plenoxel, except without the sparsity. They start by fitting degree 2 spherical harmonics (27 parameters per voxel), and then fine tune on degree 0 spherical harmonics (3 parameters per voxel, essentially representing a view-independent color value). Which would mean the actual diffusion process operates on voxels with only 4 parameters (density + rgb). I'm curious how long the voxel fitting took for each of the input images. If Plenoxel took ~10 minutes per scene, and this was trained on 15,000 chairs, then did this dataset require something like 2500 GPU-hours to generate?

[@pranavdeshpande4942]

Where is the code?

[@georgetang50]

Please release the code