Consistent Novel View Synthesis without Explicit 3D Representation

Free3D, a simple and efficient method, can achieve consistent novel view synthesis without relying on an explicit 3D representation by introducing a ray conditioning normalization layer and a pseudo-3D cross-attention module.