04 — Attention

What the model looks at

The attention map highlights the regions that pushed the “failed” score up. Warmer, more opaque areas contributed more; cooler, more transparent areas contributed less.

Each sample carries its probability, prediction and ground truth burned into the top-left corner, so the outcome is readable without interpreting the heat colours.

Where it breaks down

The same printer and scene as the previous sample three frames later, with attention heat on the bed texture and a spool holder rather than on the stray filament
good — 13.6%, wrong. This is the same print, three frames later than the sample above. The score collapsed from 69.7% to 13.6% and the heat drifted onto the bed texture and a spool holder instead of the stray filament.

Early-stage failures flip between adjacent frames like this, which is exactly why the plugin never alerts on a single frame. It waits for 3 of the last 5 to cross the threshold — see the front page overview.