Splat the Net: Radiance Fields with Splattable Neural Primitives

The paper introduces splattable neural primitives, a hybrid representation combining neural density fields with primitive-based splatting for efficient novel view synthesis. It resides in the Explicit Grid and Primitive-Based Representations leaf under Efficient Representation and Real-Time Rendering, alongside four sibling papers including Plenoxels, PlenOctrees, and point-based methods. This leaf contains five papers total within a taxonomy of fifty works, indicating a moderately populated research direction focused on balancing rendering speed with representational quality through explicit geometric structures rather than purely implicit neural networks.

The taxonomy reveals neighboring branches addressing complementary efficiency challenges: Compact Encodings and Compression focuses on model size reduction through hash encodings and pruning, while Distillation and Hybrid Representations bridges neural and light field models. The parent category Efficient Representation and Real-Time Rendering sits alongside branches handling sparse-view generalization, dynamic scenes, and specialized sensors, reflecting the field's diversification beyond foundational NeRF formulations. The scope note for this leaf explicitly excludes purely implicit MLP methods and dynamic representations, positioning the work within a cluster prioritizing fast rendering through explicit primitives like voxels, octrees, or point clouds.

Among thirty candidates examined, the analytical line integral solution shows the most substantial prior overlap, with four refutable candidates identified from ten examined for this contribution. The splattable neural primitives representation itself appears more distinctive, with zero refutable candidates among ten examined, suggesting limited direct precedent for bounded neural density fields as splatting primitives. The taxonomy framing contribution found no refutable work among ten candidates. These statistics reflect a focused semantic search rather than exhaustive coverage, indicating that while the core primitive design may be novel within this search scope, the analytical rendering technique has more established foundations in the examined literature.

Based on the limited search of thirty semantically related papers, the work appears to occupy a recognizable position within primitive-based efficiency research, with its main novelty likely concentrated in the specific formulation of neural primitives rather than the broader splatting paradigm. The analysis does not cover potential overlaps outside top-ranked semantic matches or recent concurrent work, and the refutability counts reflect only the examined candidate set, not the entire field.