This case is fine-tuned on the Objaverse dataset rendered by blender from multi-views.
I changed the parameters in the config and added another condition: height_z. Since the initial 3D video can only be guided and controled by the azimuth and polar in version sv3d_p and sv3d_u, w/o the radius. Hence, I added it as the height_z and change the .yaml
files, to satisfy the conditioner config that the value of embed_size
equals to the sum of all the embed_key.shape[-1]
.
Coming soon.