Tune the anti-vision pipeline and watch the human-vs-machine gap. Left preview animates (what a human sees); right is a single frozen frame (what a screenshot→classify bot sees).
Persistence of vision averages the per-frame noise away; the object stays readable.
Any single capture carries full-amplitude noise plus the baked perturbation and texture conflict.
The figure before temporal noise: texture conflict + fragmented contour + baked FGSM.