External baseline vs sandbox vs champion — one board

Rows per example: GT · CameraAnything (ECCV'26, ran on OUR dev clips via the ctx18 adapter) · expL_uni · expL_nk · expK (T2V-1.3B sandbox @5-6k) · expE_sync (14B champion @1500). Same clips, same frames (0/20/40/60/80), same fwd18-k metrics. 2026-07-29
modelavg seen-PSNRavg unseen-graynote
expE_sync (14B) @150013.20−22.3fleet champion
expL_nk (1.3B) @600012.88−22.1seq-ctx + narrow k-field — best sandbox line
expK (1.3B) @500012.60−13.517-ch channel + SYNC
expL_uni (1.3B) @600011.08−39.4no k-field: rich but wrong (texture soup)
CameraAnything (14B)9.75−31.9sharp imagery, invents its own room; weak trajectory obedience
TEST 111b656f
TEST 2501f488
TRAIN 00534f58
TRAIN 0003dc82
CameraAnything on the official ctx18 20-clip set: psnr_extrap 11.35 (last of 12), FID 77.3, MEt3R 0.107, rot-err 43.8° (worst) — ctx18 leaderboard. Caveat: translation similarity-scale 2.0 (their training range) may under-drive motion on our large-baseline clips; a trans_scale sweep is the fair-shake follow-up. Links: project · parade · atlas