YuNet + SFace goes 5 for 5 as faster Jev stumbles in face-finding test

The experiment grew out of a Qiita post in which Jev found a Waldo-style character in 0.4 seconds, prompting this spin-off that swapped the target for a real person's face. Rather than pitting raw models head-to-head, the author assembled four different search systems — Jev via its Box feature, Clef through coordinate choices, Claude Opus 5.5 answering from images, and OpenCV's YuNet plus SFace — and ran them on the same reference image and five stage images.

The most striking contrast is not speed but method. YuNet + SFace first detects faces, then compares each candidate to a reference image, a two-step split. The general-purpose models were asked to return a location straight from the image. The author notes this may be why the face-specific pipeline did well, while cautioning that if YuNet fails to detect a face, SFace cannot pick it. Similarly, Clef's 0/5 is tied to the coordinate-choice output format used here, not proof of weak image understanding.

Timing also deserves care: Jev's 3.81 seconds and YuNet + SFace's 9.94 seconds include API communication and model loading, so they are not pure inference speeds. The author's takeaway is that for face-finding, matching the decomposition of the problem to the tool may matter more than picking the newest general model.