Executive summary & key findings

Per-scene 3D Gaussian Splatting is superb at interpolating densely observed captures — and brittle exactly where it is asked to invent: extrapolative held-out views with big disocclusion holes, dynamic moving content, and long-horizon chained / loop-closing trajectories that accumulate drift. Diffusion-based generative NVS replaces 3D-GS in precisely those regimes. Our method — Wan2.1-VACE-14B warp+inpaint warps the input to each target pose, then lets video diffusion fill the disocclusions hole-free while staying geometrically anchored to the warped scaffold.

Every number below is read verbatim off the scored runs (eval50, Kubric-4D clean-GT, NVIDIA, SpatialVID, LH-DL3DV, WorldScore) and SLURM sacct timing — nothing is interpolated. "Ours" = vace14b; best-in-column values are noted as such.

Axis 1 · Static

Extrapolation, hole-free

DL3DV eval50 · 50 held-out clips

Warp+inpaint yields geometry-consistent extrapolation with no torn disocclusions: Ours leads generation realism at FID 56.4 (best in column).
Tied at the front on perceptual fidelity over the hole region only — PSNR-extrap 12.79 (vs GEN3C 12.78), full-frame LPIPS 0.534 (tied-best).
Strongest pose anchoring: COLMAP registration rate 0.761 (best in column) — the warped scaffold keeps generated frames self-consistent.
All methods cluster low on raw PSNR (this is generation, not reconstruction); the real separation lives in FID + geometric consistency, where Ours leads.

Axis 2 · Dynamic

Movers are hard for all

Kubric-4D clean-GT (primary) · NVIDIA · SpatialVID

Every method degrades as the viewpoint widens 30°→180°. Honest moving-object-masked mPSNR collapses to ~11–14 dB across all bins for everyone — the static background hides this in full-frame PSNR.
Ours shows the strongest pose-free multi-view consistency: MEt3R 0.014 on Kubric-4D (best; next-best 0.034) and 0.027 on NVIDIA (best, with TSED 0.80, rot 1.0°).
Camera-trajectory error stays low and flat across difficulty bins (COLMAP ATE 0.024→0.050 scene-units, 30°→180°).
Ours is absent from the pixel-PSNR-vs-angle curves by design: it does not follow the exact GT poses, so reconstruction metrics don't apply — it is judged on pose-free geometry/realism instead.

Axis 3 · Long-horizon

On-trajectory under drift

LH-DL3DV drift/loop · WorldScore (ICCV'25)

Warp-anchored control stays on-trajectory across long rollouts: drift ATE 3.48% of scene extent, rotational drift 2.5° — while single-shot baselines collapse rotationally (52–116°).
Best long-horizon realism: FID 50.9 (best in column) over 228-frame rollouts.
On WorldScore, Ours is 2nd overall (67.8), best photometric quality (88.9) and 2nd on camera controllability (78.3); NVS-Solver leads overall (70.4).

Honest tradeoffs

What Ours wins: hole-free extrapolated content, best-in-class geometric / multi-view consistency, and 2nd place on the WorldScore composite — without any per-scene optimization.
Runtime is mid-pack, not fastest. Feed-forward world models are far cheaper per clip: on long-horizon, Ours runs 2732 s/clip vs Lyra-2's 464 (5.9× slower) and MatrixGame-3's 354 (7.7× slower). The geometric-consistency win costs roughly 6–8× the inference budget of the fastest feed-forward baselines.
Dynamic movers remain unsolved for everyone. Masked mPSNR pins at ~11–14 dB across all methods and all angle bins on clean-GT Kubric-4D — reconstructing fast-moving content from a single view is an open problem, not a per-method gap.

Bottom line Warp+inpaint VACE-14B delivers hole-free, geometry-anchored novel views that stay on-trajectory where 3D-GS fails — best FID and consistency, 2nd on WorldScore — at a mid-pack inference cost (~6–8× the fastest feed-forward models), with dynamic movers still hard for the whole field.

Sources: quantitative.html tables, runs/dyn_kubric4d_metrics/*.json (per-method aggregate + by-bin), and infer_time.json (SLURM sacct s/clip). Ratios computed from long-horizon secs_per_clip. Where Ours has no value for a metric (e.g. exact-pose reconstruction PSNR), it is stated as not-applicable rather than estimated.