Rethinking Radiance Field Representations: Efficiency, Scalability, and Appearance

Radiance fields have emerged as a powerful paradigm for novel view synthesis, enabling high-fidelity scene reconstruction and rendering directly from multi-view image collections. Despite significant progress, three challenges limit their practical applicability: the computational cost of improving reconstruction quality, the lack of scalability in explicit representations under varying memory and bandwidth constraints, and the limited expressive power of commonly adopted directional appearance models. This thesis investigates a set of model-agnostic methods that address these challenges through principled reorganization of representations, computation, and directional parameterizations, without requiring fundamental changes to the underlying rendering framework. On the efficiency front, a sparse mixture-of-experts formulation is introduced in which hybrid radiance field models at different resolutions specialize in rendering different frequency components of the scene, improving reconstruction quality without proportionally increasing inference cost. On the scalability front, Gaussian primitives are organized into a progressive hierarchy based on gradient importance, enabling a single trained model to support multiple rate–distortion operating points through scalable compression and progressive decoding. This hierarchy is further extended to continuous rate adaptation, allowing smooth transitions between compression levels at arbitrary target bitrates without retraining. Finally, a new explicit representation for directional appearance is proposed, based on a differentiable partition of the sphere that combines the optimization stability of global bases with the ability to model localized high-frequency directional effects. The representation is applied to both view-dependent radiance modeling and spatially varying reflection modeling through learnable light probes. Taken together, these contributions demonstrate that efficiency, scalability, and expressive appearance modeling can be substantially advanced by rethinking how scene representations are structured, compressed, and parameterized.