The primary, confound-free dynamic anchor: Kubric-4D has perfect ground-truth poses, depth and segmentation, so there is no depth/scale/warp confound. Targets are binned by optical-axis change from the input view into 5 difficulty bins (standard GCD / AnyView protocol). The curves below show how every method degrades as the viewpoint widens — easy near-views (30°) on the left, hard back-views (120°/180°) on the right. Ours (VACE-14B) is the bold green line.
Clips per bin — 30°: n=20 · 60°: n=20 · 90°: n=20 · 120°: n=20 · 180°: n=20. Only methods/metrics with scored data are drawn; nothing is interpolated or fabricated.
Takeaway: all methods degrade with viewpoint change — the gap at 120–180° is the real test. Full-frame PSNR/LPIPS fall off monotonically as the back-views come in, and the honest moving-object-masked mPSNR collapses to ~11–14 dB across all bins (the static background hides this in full-frame numbers). Ours (VACE-14B) keeps camera-trajectory error low and roughly flat across bins (COLMAP ATE 0.024→0.050 scene-units, 30°→180°) — warp-anchored control stays on-trajectory even at the hard back-views.
x = camera-angle bin (optical-axis change vs. input view). PSNR/mPSNR higher-is-better; LPIPS & COLMAP-ATE lower-is-better. Ours (vace14b) and TrajectoryCrafter don't follow the exact GT target poses, so they have no reconstruction PSNR/mPSNR/LPIPS on this protocol (drawn only on the camera-trajectory chart, where Ours' warp-anchored control is the standout). Charts are inline SVG — fully offline, no external libraries.