A Text-Prompted Perceptual Metric Makes Image Similarity Context-Dependent

arXiv·high signal

Sheng-Yu Wang, Yotam Nitzan and Aaron Hertzmann (Adobe Research) argue in arXiv 2607.18237 that fixed perceptual metrics like LPIPS collapse a context-dependent human judgment into one number — two images can be similar in shape but different in color, and which matters depends on the task. Their metric takes a text prompt specifying the axis of comparison. For builders doing image eval or dedup, this means the similarity threshold can finally be specified in words.