Input / output pairs — the AV object-removal pipeline
Fourteen pairs drawn at random from the 1157 clips the pipeline has passed so far
on the speech-free, music-free source set. Left is the untouched source; right is the pipeline's
output — the object erased from the picture and its sound separated out of the mix, with the
rest of the soundtrack intact. Play with sound on.
5,427clips processed
1,157passed the gates
14random verified pairs shown
What runs between input and output
1 caption — Qwen3-Omni2 pick the target — GPT-4o-mini3 segment — SAM34 erase — EffectErase5 verify — SAM3 re-segment6 separate the sound — SAM-Audio, best of 10
Nothing but the source clip is given: the pipeline captions it, decides
what to remove, and does the removal. Clips that fail any gate are discarded — 79% of what it
saw — so these are drawn from the survivors.
How each pair was verified
SAM-Audio splits the soundtrack into a target (the object's sound) and a residual
(everything else); the residual is what you hear on the right. A pair is only shown if it fails in
neither direction: the residual keeps at least 30% of the source, so the scene is not wiped;
the target holds at least 20%, so something was genuinely taken out; and the target matches the
prompt better than the residual does. Both numbers appear on every card.
That two-sided test matters — the pipeline's built-in ranking maximises similarity to the
extracted target, which is maximised by dumping the whole mix into it and leaving silence behind.
Roughly one selection in four collapses that way and is excluded here.
removed: manJAVEdit-100kjavedit_1a3a8f2a76f2
1280×720 · 25 fps · 4.86s
INPUT · source
source audio
OUTPUT · man removed
after removal — the rest of the mix survives
source vs object removed
19%mask area
0.989visual removal score
53%of the mix pulled into the object
83%of the mix left behind
visualwinning prompt · seed 1132891577
removed: crowd of peopleInsAVE-80Kinsave_add_and_remove_06746
1280×704 · 23.98 fps · 3.42s
INPUT · source
source audio
OUTPUT · crowd of people removed
after removal — the rest of the mix survives
source vs object removed
17%mask area
1.000visual removal score
49%of the mix pulled into the object
84%of the mix left behind
visualwinning prompt · seed 1453635084
removed: womanJAVEdit-100kjavedit_0b4017e582e3
1280×720 · 25 fps · 4.86s
INPUT · source
source audio
OUTPUT · woman removed
after removal — the rest of the mix survives
source vs object removed
16%mask area
1.000visual removal score
73%of the mix pulled into the object
65%of the mix left behind
textwinning prompt · seed 1132891577
removed: large brown bearInsAVE-80Kinsave_general_editing_36342