Fair common-N comparison

Why this section? Different methods were scored on different numbers of clips (different N) — some runs are partial, some methods skip clips they cannot handle — so the headline aggregates elsewhere are not strictly comparable. Here we fix that per benchmark: we take the intersection of clip ids (cid) present in every scored method, then recompute each method’s mean metrics over that identical clip set and report N_common. Methods still scoring or with very few clips are marked pending and excluded from the intersection (so a half-finished run does not collapse N_common). Where a model does not produce reference-pixel metrics (it generates freely instead of following exact poses), we split into a pixel-metric group (exact-pose methods) and a pose-free group (geometry/perceptual metrics, all methods incl. Ours), each with its own N_common. All numbers below are recomputed from per-clip rows on disk — arrows mark higher-(↑) or lower-(↓) better; green = best, the highlighted row is Ours.

Kubric-4D (synthetic dynamic NVS, perfect-GT anchor)

Exact-pose methods are scored with reference pixel metrics; our model and TrajCrafter generate freely (do not follow exact 81 poses) so they have only pose-free metrics. We therefore form two common-N groups.

Pixel-metric group: N_common = 90 (intersection of 6 methods)
Methods (reference-pixel): FlexWorld, FrameCrafter, GEN3C, I2V-Base, NVS-Solver, ViewCrafter. Intersection gated on [psnr_full]; every aggregate recomputed on this identical clip set.
Method PSNR↑ SSIM↑ LPIPS↓ PSNR-dyn↑
cov ≥0/90
I2V-Base 27.48 0.7393 0.2782
FrameCrafter 26.78 0.6807 0.3183 10.98
NVS-Solver 25.21 0.7196 0.2777 14.00
ViewCrafter 25.01 0.7299 0.3256 11.52
GEN3C 19.48 0.7220 0.3419 11.75
FlexWorld 15.32 0.5904 0.4047 12.10
Pose-free group: N_common = 60 (intersection of 7 methods)
Methods (all incl. Ours): FlexWorld, FrameCrafter, GEN3C, NVS-Solver, TrajCrafter, VACE-14B (Ours), ViewCrafter. Intersection gated on [met3r, flow_warp_psnr]; every aggregate recomputed on this identical clip set.
Method MEt3R↓ FlowWarp-PSNR↑ COLMAP-ATE↓
cov ≥37/60
VACE-14B (Ours) 0.0143 44.45 0.0774
GEN3C 0.0343 37.33 1.3646
FlexWorld 0.0428 40.80 0.2454
NVS-Solver 0.0633 29.76 2.1745
TrajCrafter 0.0900 30.52 0.8654
ViewCrafter 0.1076 25.42 0.3584
FrameCrafter 0.1125 24.56 0.0003

Pending (still scoring / too few clips, excluded from the intersection so they do not shrink N_common): Shape-of-Motion (N=15), Voyager (N=5).

NVIDIA Dynamic Scenes (real dynamic, mono diagonal protocol)

Same split as Kubric-4D: pixel metrics over exact-pose methods; pose-free metrics (MEt3R, COLMAP-ATE) include our model and TrajCrafter.

Pixel-metric group: N_common = 18 (intersection of 9 methods)
Methods (reference-pixel): FlexWorld, FrameCrafter, GEN3C, I2V-Base, NVS-Solver, Shape-of-Motion, ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [psnr_full]; every aggregate recomputed on this identical clip set.
Method PSNR↑ SSIM↑ LPIPS↓ PSNR-dyn↑
cov ≥0/18
FrameCrafter 29.09 0.6780 0.1622 13.62
I2V-Base 27.27 0.6407 0.3461
NVS-Solver 26.64 0.6801 0.2892 13.85
ViewCrafter 25.80 0.6219 0.3590 11.30
GEN3C 18.77 0.5559 0.3713 13.20
ViewExtrapolator 15.52 0.3462 0.5637 12.32
FlexWorld 14.89 0.3479 0.4851 12.00
Shape-of-Motion 14.89 0.4226 0.7855 10.71
Voyager 13.91 0.2932 0.6010 11.39
Pose-free group: N_common = 18 (intersection of 10 methods)
Methods (all incl. Ours): FlexWorld, FrameCrafter, GEN3C, I2V-Base, NVS-Solver, Shape-of-Motion, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [met3r]; every aggregate recomputed on this identical clip set.
Method MEt3R↓ COLMAP-ATE↓
cov ≥10/18
VACE-14B (Ours) 0.0267 0.1330
FlexWorld 0.0673 0.5131
Voyager 0.0707 2.8937
I2V-Base 0.0977 2.9697
GEN3C 0.1050 1.5978
NVS-Solver 0.1063 0.6997
ViewCrafter 0.1123 0.3383
FrameCrafter 0.1230 0.7612
ViewExtrapolator 0.1239 0.6799
Shape-of-Motion 0.2834 0.1046

Pending (still scoring / too few clips, excluded from the intersection so they do not shrink N_common): TrajCrafter (N=14).

SpatialVid (in-the-wild dynamic camera-trajectory set)

Our model follows exact poses on a subset of clips, so it appears in the pixel-metric group on those common clips; the pose-free group spans all methods including TrajCrafter.

Pixel-metric group: N_common = 10 (intersection of 8 methods)
Methods (reference-pixel): FrameCrafter, GEN3C, I2V-Base, NVS-Solver, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [psnr_full]; every aggregate recomputed on this identical clip set.
Method PSNR↑ SSIM↑ LPIPS↓ PSNR-dyn↑
cov ≥0/10
FrameCrafter 27.22 0.7099 0.2537 16.83
I2V-Base 27.03 0.7096 0.3263
ViewCrafter 24.82 0.6481 0.3977 12.40
NVS-Solver 24.71 0.7059 0.3146 16.79
GEN3C 16.62 0.5925 0.4093 16.33
VACE-14B (Ours) 15.53 0.4318 0.5736 15.85
ViewExtrapolator 14.70 0.4599 0.5535 14.74
Voyager 13.79 0.4977 0.5456 14.97
Pose-free group: N_common = 14 (intersection of 9 methods)
Methods (all incl. Ours): FrameCrafter, GEN3C, I2V-Base, NVS-Solver, TrajCrafter, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [met3r]; every aggregate recomputed on this identical clip set.
Method MEt3R↓ COLMAP-ATE↓
cov ≥13/14
I2V-Base 0.0608 0.5303
GEN3C 0.0613 0.7709
Voyager 0.0638 0.2891
VACE-14B (Ours) 0.0711 0.1443
NVS-Solver 0.0787 0.3673
TrajCrafter 0.0898 0.4490
ViewCrafter 0.1001 0.1211
ViewExtrapolator 0.1089 0.2257
FrameCrafter 0.1416 0.3547

DL3DV eval50 (50-clip static-scene extrapolation)

All methods follow exact poses except TrajCrafter; the pixel group covers exact-pose methods (incl. Ours), the pose-free group adds TrajCrafter.

Pixel-metric group: N_common = 34 (intersection of 9 methods)
Methods (reference-pixel): FlexWorld, FrameCrafter, GEN3C, I2V-Base, NVS-Solver, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [psnr_full]; every aggregate recomputed on this identical clip set.
Method PSNR↑ SSIM↑ LPIPS↓
FrameCrafter 26.48 0.6586 0.2365
I2V-Base 24.66 0.5968 0.3598
ViewCrafter 24.61 0.6160 0.3512
NVS-Solver 23.26 0.6302 0.3726
GEN3C 14.68 0.4444 0.5896
VACE-14B (Ours) 14.48 0.4286 0.5201
FlexWorld 13.82 0.4052 0.5987
ViewExtrapolator 13.31 0.3611 0.6557
Voyager 11.54 0.3621 0.7170
Pose-free group: N_common = 32 (intersection of 10 methods)
Methods (all incl. Ours): FlexWorld, FrameCrafter, GEN3C, I2V-Base, NVS-Solver, TrajCrafter, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [met3r]; every aggregate recomputed on this identical clip set.
Method MEt3R↓ COLMAP-ATE↓
cov ≥21/32
I2V-Base 0.0529 1.9535
NVS-Solver 0.0772 1.3416
Voyager 0.1023 0.4459
FlexWorld 0.1057 0.4528
VACE-14B (Ours) 0.1172 0.4089
FrameCrafter 0.1337 0.6157
GEN3C 0.1345 0.6805
ViewCrafter 0.1382 0.1749
TrajCrafter 0.1883 0.7994
ViewExtrapolator 0.1987 0.5425

Long-Horizon DL3DV (extended-trajectory generation)

Reference pixels exist only on the GT-backed extrapolation clips (psnr_extrapGT), so the pixel group intersects over those clips; the pose-free group uses MEt3R / drift-ATE / collision over all clips. All scored methods follow exact poses here.

Reference-pixel + pose-free group: N_common = 9 (intersection of 11 methods)
Methods (reference-pixel): FlexWorld, FrameCrafter, GEN3C, Lyra2, MatrixGame3, NVS-Solver, TrajCrafter, VACE-14B (Ours), ViewCrafter, ViewExtrapolator, Voyager. Intersection gated on [psnr_extrapGT]; every aggregate recomputed on this identical clip set.
Method PSNR-extrapGT↑ SSIM-extrapGT↑ LPIPS-extrapGT↓
VACE-14B (Ours) 13.89 0.4248 0.6366
NVS-Solver 12.64 0.3949 0.7173
Lyra2 12.61 0.3708 0.7094
GEN3C 12.57 0.3795 0.6887
FlexWorld 12.55 0.4041 0.6826
FrameCrafter 12.33 0.3290 0.6099
ViewCrafter 12.11 0.3511 0.6747
Voyager 11.55 0.3758 0.7398
TrajCrafter 11.40 0.3521 0.7530
MatrixGame3 11.27 0.3556 0.7565
ViewExtrapolator 10.63 0.2602 0.7728

Pending (still scoring / too few clips, excluded from the intersection so they do not shrink N_common): VACE-14B-ft (Ours, LH-ft) (N=13).

All aggregates in every table above are recomputed on the identical intersected clip set for that group — no method is averaged over clips another method in the group did not see.