JavisDiT 0904 remove-only (fps-fixed): epoch015 vs epoch016

Run text_film_adalnsum_remove_0904_fpsfix (training job 530408). Checkpoints epoch015-global_step24432, epoch016-global_step25959. 20 validation clips (val_fpsfix_rand20_seed0.jsonl: the first 20 ids of val_rand30_seed0, with the fps-fixed target video), object removal without the reference anchor (noref_remove), cfg 1.0 / 50 steps, --per-entry-seed so both checkpoints use the same seed per clip. This page shows 8 of the 20 clips.

Standard metrics (20 clips)

FVD and FAD are population statistics and are noisy at N=20.

checkpoint	PSNR↑	SSIM↑	LPIPS↓	FVD↓	FAD↓	IS↑	KL↓	LSD↓
epoch015 (step 24432)	20.64	0.666	0.249	1085	2.29	1.67	1.854	15.54
epoch016 (step 25959)	20.75	0.664	0.264	1000	1.96	1.74	1.750	16.02

Paired per-clip change, epoch016 − epoch015 (Wilcoxon)

metric	mean Δ	ep016 better / worse	p
PSNR	+0.108	12 / 8	0.546
SSIM	-0.003	12 / 8	0.452
LPIPS	+0.015	10 / 10	0.841
KL	-0.104	11 / 9	0.388
LSD	+0.477	12 / 8	0.985

Examples (8 clips): input · GT target · epoch015 · epoch016

-OGq-n_9EYA_000000 — remove “vehicle engine”

input (source)

GT target

epoch015 (step 24432) · PSNR 16.48 · LSD 37.29

epoch016 (step 25959) · PSNR 16.36 · LSD 46.92

xbS8Hc2Ldoo — remove “helicopter”

input (source)

GT target

epoch015 (step 24432) · PSNR 23.77 · LSD 12.49

epoch016 (step 25959) · PSNR 23.84 · LSD 12.47

JfuN9jY3Psg_000010 — remove “rooster”

input (source)

GT target

epoch015 (step 24432) · PSNR 16.49 · LSD 10.24

epoch016 (step 25959) · PSNR 16.59 · LSD 10.14

6boBtYqUe9c — remove “goats”

input (source)

GT target

epoch015 (step 24432) · PSNR 20.42 · LSD 12.46

epoch016 (step 25959) · PSNR 20.62 · LSD 12.35

EjoXNd5uye0 — remove “power tool”

input (source)

GT target

epoch015 (step 24432) · PSNR 19.80 · LSD 16.69

epoch016 (step 25959) · PSNR 19.06 · LSD 15.98

z-YEPSzdgpk — remove “woodwind instrument”

input (source)

GT target

epoch015 (step 24432) · PSNR 16.53 · LSD 11.48

epoch016 (step 25959) · PSNR 25.46 · LSD 9.95

rYmGpcue7cc_000012 — remove “trumpet”

input (source)

GT target

epoch015 (step 24432) · PSNR 20.71 · LSD 35.52

epoch016 (step 25959) · PSNR 18.93 · LSD 34.91

XkM6ak6DVY8 — remove “bowed string instrument”

input (source)

GT target

epoch015 (step 24432) · PSNR 20.33 · LSD 13.41

epoch016 (step 25959) · PSNR 20.73 · LSD 13.99