Adobe Research turns image similarity into a text prompt

Adobe Research is making the case that image similarity metrics have been solving the wrong problem. In a paper posted to arXiv, Wang, Nitzan and Hertzmann argue that fixed metrics like LPIPS take a judgment that depends on context and compress it into one number. Two images can match in shape and diverge completely in color, and whether that counts as similar depends on what you're checking for, not on what a pretrained network decides.

Their fix is a metric that takes a text prompt specifying the comparison axis. Tell it you care about color, it scores color. Tell it you care about shape, it scores shape instead.

That's a real change for anyone running image eval or dedup pipelines, where an LPIPS threshold usually gets tuned by trial and error until it stops throwing false positives. A text-specified axis turns that tuning knob into a sentence you can read back and check later, instead of a number nobody can explain.

The paper doesn't say how the metric holds up across thousands of comparisons with dozens of different axes, or whether small changes in prompt wording shift the score. If every comparison needs its own carefully worded prompt, this trades a threshold you tune once for a prompt you have to get right every time. That's not obviously less work, it's just work that happens earlier and looks more legible.