V2A azimuth scores — 4-ball set
Does the generated stereo image follow the ball on screen? 4 ball types × 8 trajectories, rendered as raw Kubric and as photoreal VACE re-renders, sound generated by LTX-2, Cosmos3 and PrismAudio with several seeds each. Every number below is one generated clip — click it to watch that exact video with its per-event measurements.
Two text prompts, same negative prompt and same seeds, so the comparison is paired:
neutral — A ball bouncing on the ground. Clear rhythmic bounce impacts, each bounce a short percussive sound, decaying between hits, quiet ambience in between.
and ball-named (added 2026-10-05, 10 seeds per model) — the same text with A ball replaced by
A ping pong ball / A tennis ball / A soccer ball / A basketball.
Per clip. Impact onsets are detected in the generated audio. At each onset,
ILD = 10·log10(EL/ER) where E is the mean-square of the 1–4 kHz band over
[onset−2 ms, onset+3 ms] (+ = left louder). The ball's true azimuth at that moment comes from the renderer's ground truth.
Saz = −Spearman(azimuth, ILD): +1 = louder on the side where the ball is, perfectly ordered; 0 = no relation. Clips that are dual-mono (L/R correlation > 0.999 and no onset with |ILD| ≥ 0.5 dB) or have fewer than 3 usable onsets have no ρ and count as 0 in the summary. (The |ILD| clause was added 2026-10-05 so a weak but audible pure-amplitude pan is not discarded; it changed 9 of 1096 neutral-prompt clips and no headline number.)
Top-K. For each video (ball × trajectory): S@K = best of the K seeds, S̄ = mean over seeds, averaged over the 4 balls. S@3 = expected best of 3 seeds (comparable across models with different K). Grey numbers = 95th percentile of the same statistic when ILD is shuffled within each clip — the score reachable by chance. A max over noisy seeds is far above 0 by luck alone; read every number against its grey one.
Summary
Bold = above chance. Click a row to open its per-clip table.
Neutral vs ball-named prompt
Per seed, Tp = −(mean ρ on l2r + mean ρ on r2l)/2 (drift-free tracking score, +1 = perfect). Paired = same model, style and seed under both prompts. A positive ΔTp means the named prompt tracked the ball better.
Wilcoxon signed-rank on the paired per-seed ΔTp. Few neutral seeds exist for some rows (VACE: LTX-2 6, Cosmos3 3; PrismAudio 3), so those comparisons are weak. Dual-mono = share of that model·style's clips gated as dual-mono, averaged over seeds.
Per-clip scores
Shift/Ctrl/Cmd-click a score to open the raw .mp4 in a new tab. All files:
dataset (balls_v2a/generations/gen/ neutral, gen_named/ ball-named), raw tables
per_clip_with_urls.csv, per_event.csv, topk_azimuth.csv.
Cosmos3 and PrismAudio were scored from their lossless WAV output; the linked .mp4 carries an AAC copy of the same audio
(AAC changes 1–4 kHz ILD by a median 0.15 dB).