Why terrain elevation maps need different neural compression than photos

Terrain elevation data can be compressed as continuous neural functions, but existing amortized methods designed for images don't work well for heightfields. Researchers benchmarked three approaches on 1 m/pixel terrain, then proposed HUVR+SIREN—a hypernetwork that swaps the decoder for one with analytical derivatives. It beats prior methods on height and slope fidelity while using the same storage and tolerating extreme post-training quantization. The work clarifies which design choices actually matter for terrain and suggests the bottleneck lies in the shared network's architecture, not the per-tile encoding.