Closed aradhyamathur closed 1 year ago
I wanted to test with vanilla nerf implementation like stable diffusion, the outputs there are 3 channel, so wanted to know what change could be made for the same.
The four channels are for the latent features of Stable Diffusion. It’s not RGBa. The density is controlled by another small 1-channel voxel.
Oh I see, so it's essentially predicting the latents which are then denoised by the SD network to produce the RGB while density is predicted from the other module. Is that correct ?
That’s right.
Thanks for the clarification.
Hi, I have a query pertaining to the voxnerf implementation, the features generated 4 channels, whereas the density feature is computed separately, does the last channel correspond to alpha ?