SmolVLM silently mangled images for ages as bf16 patch coordinates landed in wrong buckets

LysandreJik · x · 2026-10-22

The SmolVLM team shares a subtle bug: in their CI, patch coordinates cast to bf16 lost precision, so patches landed in the wrong position buckets — the model was effectively seeing a misaligned mosaic of the image (see the Statue of Liberty screenshot). Generation was only slightly off, so no agent or test caught it until a human said "hey, isn't that weird?" A good reminder that numerical-precision bugs in vision pipelines produce subtle anomalies only humans tend to notice.

Original post →

More from Fun

Fun channel →