← All models · ViT

use case · wild

Wild: fool the model

A classifier is only as good as what it can see. Zoom in, rotate, or slap an occlusion box over the subject, then re-classify and watch the top guess — and its confidence — fall apart. It's the fastest way to build intuition for where a vision model breaks.

Two cats A bedroom A riverside
…or drop your own image

Tip: classify once with everything at zero to set a baseline, then push one slider at a time. A big drop in the top probability (or a jump in entropy toward 1) means the model has lost the plot.

← Back to ViT · Next: Multi-model →