Class Conditioned Mesh Generation
Method
Our 3D representation is based on deep marching tetrahedra (DMTet) method. And we use a latent diffusion model for generation.
To embed the class condition, we use concatenation of the class embedding and the latent code as the input of each block of unet. We also use cross attention layer to fuse the class embedding and the latent code.
Classifier guidance is used for the generation process.
Results
Text-to-Texture Generation
Method
We distill the knowledge from 2D diffusion model to optimize the trainable texture.