NeRF (Neural Radiance Fields) models integrate geometry and appearance through volume rendering. As a result, using NeRF for 3D modeling can be less effective when it comes to capturing both the intricate details of a surface as well as its material and texture.
DMTet fine-tunning disentangles the learning of geometry and appearance models, such that both a fine surface and a rich material/texture can be generated. To enable such a disentangled learning, a hybrid scene representation of DMTet is used.
The DMTet model maintains a deformable tetrahedral grid that encodes a discretized signed distance function and a differentiable marching tetrahedra layer that converts the implicit signed distance representation to the explicit surface mesh representation.