LiTo: Surface Light Field Tokenization

The paper proposes a unified 3D latent representation that jointly encodes object geometry and view-dependent appearance by treating RGB-depth images as samples of a surface light field. It resides in the Neural Surface Light Field Models leaf, which contains six papers including the original work. This leaf sits within the broader Surface Light Field Representation and Modeling branch, indicating a moderately populated research direction focused on deep learning approaches to surface light field encoding. The taxonomy shows this is an active area with multiple competing neural methods.

The taxonomy reveals neighboring leaves addressing complementary aspects: Compression and Compact Representation focuses on traditional factorization techniques, while Geometry-Dependent Modeling and Optimization emphasizes explicit geometry with robustness to errors. Reflectance and Material Modeling tackles BRDF decomposition within surface light field frameworks. The paper's approach bridges these directions by jointly modeling geometry and appearance in a learned latent space, diverging from methods that treat these components separately. The taxonomy's scope notes clarify that this leaf excludes traditional compression methods and explicit geometric approaches, positioning the work squarely in the neural implicit representation paradigm.

Among sixteen candidates examined across three contributions, no clearly refuting prior work was identified. The core 3D latent representation examined ten candidates with zero refutations, suggesting limited direct overlap in the specific formulation of joint geometry-appearance encoding via surface light field subsampling. The training framework contribution examined four candidates without refutation, while the latent flow matching component examined two candidates. This limited search scope—sixteen papers from semantic search and citation expansion—means the analysis captures nearby work but cannot claim exhaustive coverage of all potentially relevant neural surface light field methods or generative 3D models.

Based on the examined candidates, the work appears to occupy a distinct position within neural surface light field modeling, particularly in its joint latent encoding strategy and flow-based generative component. However, the analysis is constrained by the top-K semantic search scope and does not cover the full breadth of recent neural 3D generation or view synthesis literature. The taxonomy context suggests the field is moderately crowded with competing neural approaches, warranting careful positioning against sibling methods in the same leaf.