Large view synthesis models (LVSMs) reconstruct scenes from sparse views and render novel views quickly, but they typically use a fixed set of scene anchors, leading to token redundancy and unnecessary storage and rendering costs. GLoD-LVSM addresses this by learning a continuous field that initializes a flexible number of anchors to match a desired budget. A condenser then aggregates scene information into these anchors, with both the field and condenser trained end-to-end.

The condenser also infers reference-token and anchor geometry, using geometry-grounded attention to guide compression. This improves rendering quality and enables frustum culling during decoding, so unseen regions can be skipped. Compared with existing LVSM scene-compression baselines, the authors report the best trade-off between rendering quality and scene token count.

Against the uncompressed base model, GLoD-LVSM reduces the scene-token count by 33.4% and accelerates rendering by 1.23x with no quality loss. Under a relaxed quality-loss budget, the reduction grows to 85.5% with a 3.00x rendering speedup. The authors also observe that view-synthesis supervision produces emergent scene depth and self-organized anchors that align with surfaces and adapt their density to local texture complexity, and they show the representation can support scene-semantic understanding beyond rendering.