Accurate evaluation of spatial simulation for camera‑controllable generation

Develop rigorous and unambiguous evaluation methodologies and benchmarks to accurately assess spatial simulation in camera‑controllable image generation, addressing the ambiguity of calibration‑based evaluators when generated images exhibit subtle geometric differences so that geometric consistency can be measured precisely.

Background

The paper currently evaluates camera‑controllable image generation by estimating pixel‑wise camera maps using an offline vision‑based calibration method and then computing angular errors. The authors note that, while this is the best available practice, the reported calibration errors can be ambiguous, especially when spatial differences between generated images are subtle.

They emphasize that precise evaluation of spatial simulation is essential to advance camera‑controllable generation and suggest using stronger camera understanding models as evaluators and designing benchmarks that better capture geometric consistency.