Perceptual Metrics for Evaluating Spatial and Temporal Consistency in Edited Visual Media

Abstract

Contemporary image and video editing systems enable diverse spatially localized and temporally extended modifications, including object insertion, background replacement, relighting, retiming, and generative synthesis. As these operations become more accessible and numerous, the assessment of visual quality requires metrics that explicitly account for spatial and temporal consistency rather than only global distortion with respect to a nominal reference. Spatial consistency concerns how edited regions integrate with surrounding content in terms of geometry, appearance, and semantics, whereas temporal consistency concerns how modifications evolve over time without causing flicker, motion discontinuities, or structural drift. Human observers judge these properties jointly, integrating local evidence across space and time under constraints of visual attention and memory, so computational metrics that ignore these interactions may correlate weakly with perceived plausibility. This text discusses the formulation of perceptual metrics designed to evaluate spatial and temporal consistency in edited visual media, emphasizing representations that combine low-level gradients, mid-level structures, and high-level learned features. It examines how spatiotemporal derivatives, graph-based regularity measures, and probabilistic models of human preference can be used to define differentiable objective functions suitable both for evaluation and for guiding optimization-based editing algorithms. It also considers numerical issues associated with large-scale video data, including sampling strategies, stability of gradient-based optimization, and computational trade-offs. Finally, it outlines experimental protocols for benchmarking such metrics against human judgments, with attention to the diversity of editing operations and viewing conditions, and identifies open questions in aligning metric predictions with human perception of edited visual media.