V2A Benchmark v0
Silent videos with known 3D geometry, and the audio that video-to-audio models generate for them. The model only sees the
video. One tab per cue.
Listen with headphones: a correct model puts each impact on the side of the screen where it happens.
Tool strikes: a hammer, a spoon or a cleaver strikes five targets at fixed azimuths (about −30°, −16°, 0°,
+16°, +30°; 2 strikes each) in 9 orders. Bouncing balls: five real ball types at true size and bounciness, each on 7 paths
(left to right, right to left, static left / centre / right, two diagonals). "Depth map" shows the scene depth (brighter = closer; for
photoreal sets this is the video model's control input).
Category
Version
Set
Sound
Seed