This metric evaluates whether lighting, chroma, and texture at a revisited viewpoint return to how they first appeared. It discards high-frequency content on purpose, targeting the failure where geometry is still there but the scene is relit or restyled. Spatial CIELAB means ignore object outlines, so this metric does not replace the DINOv2 / SSIM reading in memory_revisit_fidelity.
It uses the same pose-paired frames as revisit fidelity. A failed memory_return_gate does not block scoring; it uses the same nearest-pose fallback as fidelity when strict revisit pairs are absent, and records gate status and pair source. The revisit frame is resized to the reference size and converted to OpenCV 8-bit CIELAB. Lightness is the absolute difference of mean L restored to [0,100]; chroma is the Euclidean distance between mean a,b after removing the 128 offset:
The two deltas map to exponentially decaying similarities; the texture term is the same transform of the VGG Gram distance dG (the descriptor used by style_consistency):
Defaults are τL=6, τC=8, and τG=0.002. When all three terms exist they are averaged; a missing VGG checkpoint drops the texture term and averages lightness and chroma only. The video score is the mean over the selected K pairs:
In WorldAtlas this algorithm is implemented as the official metric memory_rendering_recovery (cielab_gram backend). Under the default normalization config, raw and normalized outputs lie in [0,1] with higher is better. On memory_loop it is official whenever the selected pairs can be scored, including after a failed gate; the texture term may degrade on its own.
