Ours (VACE-14B) does not follow the exact GT target poses of the GCD / AnyView protocol, so it has no reconstruction PSNR / mPSNR / LPIPS on the pixel-aligned curves — that is why it is absent from the headline difficulty chart. The metrics below are pose-free: they score the generated video on its own terms rather than against a pose-locked GT frame. Multi-view consistency (MEt3R, via DUSt3R) and temporal flow-warp consistency (optical-flow warp PSNR / EPE) are computed for Ours, so Ours finally shows up across the full 30→180° difficulty sweep alongside every pose-following baseline. Each method's curve is its own rows[] aggregated per angle bin (mean, nulls dropped) — the same recipe for all methods, so the comparison is apples-to-apples.
Pose-free clips scored for Ours per bin — 30°: n=18 · 60°: n=18 · 90°: n=18 · 120°: n=19 · 180°: n=18. Only methods/metrics with scored data are drawn; nothing is interpolated or fabricated.
Takeaway: on the pose-free axes Ours holds up flat across the whole difficulty sweep: multi-view consistency MEt3R stays low (0.011→0.016, 30°→180°) and temporal flow-warp PSNR stays high (~44.5→43.8 dB) with mean flow EPE near-constant (~0.251→0.279 px) — the warp-anchored generation does not fall apart as the camera swings to the hard back-views, whereas several pose-following baselines lose multi-view / temporal consistency at 120–180°. These axes do not reward GT-pose matching, so they are the fair place to read Ours against the field.
x = camera-angle bin (optical-axis change vs. input view). MEt3R & flow EPE lower-is-better; flow-warp PSNR & TSED inlier-fraction higher-is-better. FID is reported only as a single per-method aggregate, not per bin (no per-bin FID exists in the data, and Ours has no aggregate FID either), so we chart flow EPE in its place rather than fabricate a per-bin FID curve. Every method's curve is computed by aggregating its own rows[] by angle bin (mean, nulls dropped); i2v-base has no flow metrics and is omitted. Charts are inline SVG — fully offline, no external libraries.