WorldScore benchmark (ICCV 2025) · separate track

WorldScore metrics — world-generation quality on WorldScore's own camera tasks

WorldScore is a different benchmark from the PermaVid revisit set. It runs on its own scenes (8 static camera tasks: push-in / pull-out / move / orbit / pan) and scores an overall WorldScore plus 7 dimensions (no per-frame GT — these are learned/perceptual metrics). 9 of our methods were registered as generators and scored here.

⚠ Scope notes: (1) evaluated on 8 examples (one per camera layout), a subset of the full 2000-task leaderboard. (2) The VACE shown is the DL3DV long-horizon driver checkpoint — our UE/PermaVid finetune (ours_vace) was not run on WorldScore. (3) The dedicated finetuned run vace14b_ft_lh@4000 returned all-zeros (eval did not complete) → excluded pending a rerun.
overallWorldScore leaderboard (this subset)
metricsAll 7 dimensions per method
#methodWorldScore↑camera controlobject controlcontent align3D consist.photometricstyle consist.subjectiven
1NVS-Solver 🏆70.4 69.4 65.6 39.3 91.6 81.2 94.2 51.6 8
2VACE-14B (DL3DV-lh ckpt)67.8 78.3 56.2 33.0 85.4 88.9 93.2 39.8 8
3FlexWorld65.1 79.5 34.4 27.3 88.3 89.6 88.1 48.4 8
4Voyager63.3 75.0 50.0 33.9 80.5 83.5 89.2 31.1 8
5ViewExtrapolator61.1 76.6 59.4 32.7 79.5 56.0 87.0 36.7 8
6TrajectoryCrafter56.3 33.2 37.5 27.2 77.2 82.1 88.0 48.8 8
7GEN3C44.4 53.4 40.6 11.1 53.2 38.8 84.1 29.5 8
8FrameCrafter41.1 51.0 46.9 15.2 37.2 0.0 91.5 46.2 8
9ViewCrafter32.3 20.9 37.5 22.2 32.4 0.0 56.0 57.0 8

Cells shaded green (high) → red (low) per value (0–100). NVS-Solver tops this subset (70.4); VACE-14B (DL3DV-lh) is 2nd (67.8). These are perceptual / camera-control / consistency scores on WorldScore scenes — orthogonal to the PermaVid revisit PSNR.