Traditionally, 3D reconstruction has been addressed by Multi-View Stereo (MVS), which has achieved great geometric accuracy, but has treated "texture-mapping" the inferred surfaces as an afterthought. By focusing solely on geometric criteria, 3D representations extracted using MVS methods have yielded unsatisfactory Novel View Synthesis (NVS) results. On the other hand, recent methods based on radiance fields have generated compelling synthesized views, but their representations is of lower geometric quality because they rely primarily, or even exclusively, on photometric losses. In this talk, I will present ongoing research on 3D representation and reconstruction mechanisms that aim to infer 3D models that are both geometrically accurate and able to synthesize novel views well. I will present two approaches based on explicit radiance field representations and inference-time optimization that introduce ways for the primitives to interact in 3D, complementing the current viewpoint-based losses. We named the first approach Radiant Triangle Soup (RTS) due to the representation which is in the form of translucent triangle primitives. The key novelty of RTS is the introduction of soft connectivity forces between triangles during optimization, encouraging explicit, but soft, surface continuity in 3D. The second approach is named TVGS for Tensor-Voting based Gaussian Splatting. TVGS also enables direct communication among primitives to enhance the geometric structures they form in 3D. This is accomplished by Tensor Voting, which was originally designed to infer structures from noisy inputs and has been adapted here to provide supervision during inference-time optimization, leading to more accurate scene geometry. The 3D losses introduced in both approaches do not require rendering and can be combined with essentially all losses previously reported in the literature.