0902_ropealign vs sharedt_ropealign · training samples step 6000 & 6500
Validation-callback samples (samples/step_N), all 32 clips per step. Both runs sample the same 32 clips at a given step (src/target byte-identical) → paired comparison. Step 6000 and 6500 share 0 clips → across steps only the margin vs each step's own pass-through is comparable.
FAD: login-node CPU VGGish, single-thread BLAS (scipy sqrtm = symmetric form). KL/IS/LSD(aldm): audioldm_eval MERT, LSD offset 0, KL_FIX. Wave PSNR/LSD: per-clip waveform vs target. Video: per-clip PSNR/LPIPS vs target. n=32 → KL/IS are near-degenerate.
Summary
Green = better of the two ropealign runs at that step. Italic = pass-through reference.
step
arm
FAD ↓
KL ↓
IS ↑
LSD (aldm) ↓
wave PSNR ↑
wave LSD ↓
video PSNR ↑
video LPIPS ↓
6000
pass-through (src)
5.74
0.0007
1.011
14.71
20.78
10.87
16.57
0.282
6000
0902_ropealign
12.15
0.0021
1.009
16.21
20.17
10.69
18.25
0.246
6000
sharedt_ropealign
13.22
0.0022
1.010
17.32
21.29
9.79
17.67
0.251
6500
pass-through (src)
4.63
0.0009
1.012
12.98
19.48
9.45
16.64
0.315
6500
0902_ropealign
7.28
0.0011
1.011
13.06
17.77
9.33
17.99
0.288
6500
sharedt_ropealign
8.19
0.0018
1.011
14.71
18.94
10.04
17.69
0.291
Tests
Paired Wilcoxon within a step; Mann-Whitney on (gen − pass) margins across steps. Yellow = p<0.05.
comparison
metric
Δ
better
p
step 6000: sharedt − rope (paired)
aPSNR
+1.116
sharedt
0.0708
step 6000: sharedt − rope (paired)
aLSD
-0.898
sharedt
0.0088
step 6000: sharedt − rope (paired)
vPSNR
-0.580
rope
5.97e-06
step 6000: sharedt − rope (paired)
vLPIPS
+0.004
rope
0.00571
step 6500: sharedt − rope (paired)
aPSNR
+1.171
sharedt
0.00934
step 6500: sharedt − rope (paired)
aLSD
+0.704
rope
0.331
step 6500: sharedt − rope (paired)
vPSNR
-0.294
rope
0.0521
step 6500: sharedt − rope (paired)
vLPIPS
+0.003
rope
0.106
0902_ropealign: 6500 − 6000 (margin vs pass, unpaired)
aPSNR
-1.112
6000
0.251
0902_ropealign: 6500 − 6000 (margin vs pass, unpaired)
aLSD
+0.066
6000
0.909
0902_ropealign: 6500 − 6000 (margin vs pass, unpaired)
vPSNR
-0.335
6000
0.432
0902_ropealign: 6500 − 6000 (margin vs pass, unpaired)
vLPIPS
+0.009
6000
0.245
sharedt_ropealign: 6500 − 6000 (margin vs pass, unpaired)
aPSNR
-1.057
6000
0.984
sharedt_ropealign: 6500 − 6000 (margin vs pass, unpaired)
aLSD
+1.668
6000
0.0869
sharedt_ropealign: 6500 − 6000 (margin vs pass, unpaired)
vPSNR
-0.049
6000
0.624
sharedt_ropealign: 6500 − 6000 (margin vs pass, unpaired)