Eve library

Frame artifacts: one detector, four disqualifications (2026-09-10)

Guide

Revision 1

On this page

Evidence for 182.C.19.05, which names seven artifacts: malformed teeth/tongue/lips, jaw discontinuity, asymmetry spikes, face boundary tearing, extra limbs/digits, fused props, and impossible object contact.

One of the seven is now detected. Four candidate approaches were measured and all four failed, and the failures are the more useful half of this page — somebody will otherwise build them again.

Fixture#

The committed three-panel probe ../2026-09-10-face-mesh-artifact-probe/clean-vs-doubled-mouth-vs-tear.png: the same real face, clean, with a doubled mouth, and with a horizontal band displaced (boundary tearing). A fourth panel, a vertical tear, is built from the clean one the same way, to validate the column axis against a fault known to be there.

What works: line continuity#

A tear is a displacement — a band shifted, so at its edges one row no longer continues into the next. Correlation between adjacent rows measures exactly that.

Panel worst row corr rows < 0.7 worst col corr cols < 0.7
clean 0.796 0 0.810 0
doubled mouth 0.796 0 0.810 0
torn (horizontal) 0.292 2 0.833 0
torn (vertical) 0.558 1 0.224 2

It also names where: row 168 for the horizontal tear and column 170 for the vertical one, both the displaced band's edge. And the axis with the lower minimum names the tear's orientation — a vertical tear disturbs some rows too (0.558), but far less than it disturbs columns.

A 0.7 threshold sits between 0.796 and 0.292 with room on both sides. Two discontinuous lines are required rather than one, because a displaced band has two edges and a single bad line is as likely to be a hard cut inside the frame — a title card edge, a split screen.

What does not work#

Segmentation consistency — the item's own suggested candidate#

Panel face-skin components largest share holes boundary roughness
clean 2 (1 over 100 px) 1.000 0 19.28
doubled mouth 1 1.000 0 20.26
torn 1 1.000 0 25.63

Every structural measure is identical. Roughness moves 33% for the tear and only 5% for the doubled mouth — inside the difference between two ordinary frames.

Copy-move block matching#

The idea that a doubled mouth is a pasted copy. 24×24 blocks, stride 8, flat blocks skipped, pairs at least 40 px apart:

Panel best distant correlation best pair pairs > 0.99
clean 0.9709 (140,156) ↔ (188,260) 0
doubled mouth 0.9709 (140,156) ↔ (188,260) 0
torn 0.9707 (140,156) ↔ (188,260) 0

Identical on all three, on the same coordinate pair. The doubled mouth is not a pixel copy, so block matching cannot see it.

Dense face-mesh asymmetry and lip curvature#

Disqualified earlier the same day in ../2026-09-10-face-mesh-artifact-probe/: the mesh scored the doubled mouth as more regular than the clean face.

Still undetected#

Artifact Status
malformed teeth/tongue/lips four approaches measured and failed; remaining candidate is a trained detector (182.C.33)
jaw discontinuity no fixture; a hard displacement would show here, a soft malformation would not
asymmetry spikes mesh asymmetry disqualified
extra limbs/digits hand keypoint models emit 21 keypoints unconditionally
fused props, impossible object contact needs scene understanding, which nothing staged here provides

Reproducing#

text
python tools/human-video/evaluator/frame_continuity.py --image <frame.png>

Per-item status this evidence supports#

Item Was Is
182.C.19.05 not-started blocked-corpus: 1 of 7 detected, 4 approaches disqualified, malformed mouth needs 182.C.33