Quantitative Benchmarks

Arrows show metric polarity: ↑ higher-is-better, ↓ lower-is-better. Bold green = best in column. Highlighted row = Ours (VACE-14B). — = no value; running = method has outputs but the metric is not yet computed. Every cell is read verbatim from bench_data.json — no value is invented.

1 · Static novel-view extrapolation — DL3DV (50 held-out clips)

PSNR / SSIM / LPIPS are scored over the hole / extrapolated region only (eval_metrics psnr_extrap etc.) — the unobserved pixels a method must hallucinate. Full-frame variants shown as secondary where available. Geometry: TSED (SfM 2-view inlier frac), MEt3R (multi-view consistency), COLMAP registration rate & rotation error.

Method	Hole / extrapolated region (primary)			Full frame (secondary)			Generation realism		Geometry / consistency			n	infer s/clip↓
Method	PSNR↑	SSIM↑	LPIPS↓	PSNR↑	SSIM↑	LPIPS↓	FID↓	TSED↑	MEt3R↓	reg↑	rot°↓	n	infer s/clip↓
Ours · Wan2.1-VACE-14B OURS	12.79	0.374	0.605	14.06	0.411	0.534	56.4	0.498	0.135	0.761	17.7	50	758 (9.36/f)
GEN3C · Cosmos-7B	12.78	0.385	0.661	14.25	0.425	0.605	running	0.258	0.135	0.302	68.6	50	498 (6.14/f)
Voyager · HunyuanVideo	11.05	0.338	0.741	11.32	0.350	0.725	136.3	0.179	0.109	0.341	93.7	49	1086 (13.41/f)
FlexWorld	12.88	0.373	0.657	13.54	0.389	0.613	67.2	0.266	0.121	0.631	24.1	50	259 (3.20/f)
ViewCrafter	11.83	0.318	0.627	—	—	—	running	0.582	0.144	0.654	4.2	36	264 (3.26/f)
FrameCrafter · Wan2.1-14B	14.38	0.377	0.448	—	—	—	running	0.639	0.150	0.756	16.1	50	2413 (29.80/f)
NVS-Solver	12.98	0.380	0.641	—	—	—	running	0.483	0.088	0.776	43.9	45	122 (1.51/f)
ViewExtrapolator	11.23	0.291	0.742	12.96	0.341	0.675	116.6	0.205	0.221	0.269	80.1	50	131 (1.61/f)
TrajectoryCrafter	running	running	running	—	—	—	87.1	0.108	0.193	0.481	134.7	0	610 (12.46/f)

Full-frame PSNR/SSIM/LPIPS are present only for methods whose harness emitted them (vace14b, gen3c, flexworld, voyager, viewextrapolator); others show —. All methods cluster low on raw PSNR (this is generation, not reconstruction); separation lives in perceptual LPIPS, FID and geometric consistency. Ours leads FID (56.4) and is essentially tied-best on full-frame LPIPS (0.535).

2 · Dynamic-scene novel-view synthesis

The confound-free primary anchor is Kubric-4D clean-GT (perfect GT poses/depth/segmentation — the headline angle-binned difficulty curve is up top). NVIDIA (dynamic-region masked) is a secondary real-dynamic check — its 12-camera static rig leaks multi-view cues (the “teleporting camera”). DyCheck (iPhone portrait) was dropped: its portrait aspect ratio doesn't fit the landscape pipeline.

2a · NVIDIA Dynamic Scenes SECONDARY

24 clips, 12-camera static rig. Masked = dynamic-region.

Method	★ Masked — dynamic-region (primary)			Full frame (secondary)			Realism		Consist.	n	m·n	infer s/clip↓
Method	mPSNR↑	mSSIM↑	mLPIPS↓	PSNR↑	SSIM↑	LPIPS↓	FID↓	TSED↑	MEt3R↓	n	m·n	infer s/clip↓
Ours · Wan2.1-VACE-14B OURS	running	running	running	running	running	running	205.3	0.800	0.027	0	0	795 (9.81/f)
GEN3C · Cosmos-7B	13.80	0.338	0.069	19.53	0.641	0.275	71.1	0.139	0.102	21	21	779 (32.46/f)
Voyager · HunyuanVideo	11.39	0.236	0.081	14.32	0.296	0.539	187.6	0.039	0.071	18	18	2395 (99.79/f)
FlexWorld	12.00	0.251	0.076	14.73	0.345	0.470	184.7	0.673	0.067	18	18	663 (27.64/f)
ViewCrafter	12.26	0.262	0.090	12.14	0.299	0.701	91.2	0.263	0.107	21	21	505 (21.03/f)
FrameCrafter · Wan2.1-14B	13.96	0.270	0.067	17.50	0.378	0.334	69.7	0.296	0.124	21	21	742 (30.89/f)
NVS-Solver	13.85	0.335	0.076	16.71	0.392	0.557	82.8	0.232	0.106	18	18	318 (13.24/f)
ViewExtrapolator	12.32	0.263	0.079	15.28	0.340	0.538	213.9	0.114	0.124	18	18	144 (5.99/f)
TrajectoryCrafter	running	running	running	running	running	running	104.8	0.079	0.083	0	0	514 (10.50/f)
Shape-of-Motion · 4D-GS	11.24	0.296	0.096	13.26	0.412	0.796	355.2	0.228	0.292	21	21	—

⚠ NVIDIA Dynamic Scenes uses a 12-camera static rig (high effective multi-view factor — the “teleporting camera” flagged by DyCheck’s Reality Check). Treat as SECONDARY; the confound-free primary dynamic anchor is Kubric-4D clean-GT (perfect ground-truth, no warp confound — see the difficulty curve at the top). Note vace14b currently has n=0 / null masked recon on NVIDIA (warp caveat) — only geometry metrics (TSED 0.80, rot 1.0°, MEt3R 0.027, all best-in-col) are populated; its recon cells render as running.

2b · Kubric-4D (synthetic dynamic — summary) PRIMARY STANDARD GCD/AnyView

Aggregate across all angle bins (100 clips total). PERFECT-GT — no depth/scale/warp confound (the confound-free primary dynamic anchor). The headline angle-binned difficulty curve is at the top of the page; the per-bin table is Section 3 below. Ours runs are still being scored (recon n=0 → PSNR/SSIM/LPIPS show running).

Method	PSNR↑	SSIM↑	LPIPS↓	FID↓	TSED↑	MEt3R↓	reg↑	rot°↓	n	infer s/clip↓
Ours · Wan2.1-VACE-14B OURS	running	running	running	running	0.094	0.014	0.145	144.8	0	289 (3.57/f)
GEN3C · Cosmos-7B	13.68	0.540	0.537	71.5	0.101	0.034	0.192	120.1	100	503 (20.96/f)
Voyager · HunyuanVideo	11.89	0.489	0.568	241.2	0.085	0.056	0.758	113.0	5	1901 (79.22/f)
FlexWorld	12.47	0.422	0.563	171.8	0.109	0.042	0.350	132.6	100	459 (19.12/f)
ViewCrafter	12.34	0.505	0.609	78.0	0.025	0.110	0.363	146.8	100	226 (9.41/f)
FrameCrafter · Wan2.1-14B	12.44	0.371	0.631	107.5	0.018	0.114	0.204	141.6	100	866 (36.08/f)
NVS-Solver	14.52	0.479	0.522	81.8	0.065	0.063	0.580	135.0	100	230 (9.59/f)
TrajectoryCrafter	running	running	running	123.9	0.052	0.089	0.579	141.9	0	336 (6.86/f)
Shape-of-Motion · 4D-GS	5.38	0.370	0.855	336.8	0.035	0.064	0.017	131.2	15	1487 (61.98/f ·ps)

2c · SpatialVID (in-the-wild dynamic)

30 clips (group_0001).

Method	PSNR↑	SSIM↑	LPIPS↓	FID↓	TSED↑	MEt3R↓	reg↑	rot°↓	n	infer s/clip↓
Ours · Wan2.1-VACE-14B OURS	14.20	0.349	0.669	86.2	0.897	0.072	0.909	23.2	16	389 (4.81/f)
GEN3C · Cosmos-7B	14.91	0.536	0.479	running	0.765	0.058	0.771	89.3	20	455 (5.68/f)
Voyager · HunyuanVideo	12.57	0.425	0.593	69.7	0.666	0.065	0.914	76.5	30	1103 (13.79/f)
ViewCrafter	11.43	0.346	0.711	running	0.647	0.100	0.466	47.1	29	299 (3.73/f)
FrameCrafter · Wan2.1-14B	14.99	0.442	0.475	running	0.633	0.138	0.807	102.7	27	2262 (28.27/f)
NVS-Solver	14.34	0.462	0.561	33.5	0.664	0.079	0.637	69.3	30	150 (1.88/f)
ViewExtrapolator	13.02	0.390	0.641	85.9	0.674	0.111	0.625	65.3	30	129 (1.61/f)
TrajectoryCrafter	running	running	running	53.3	0.517	0.090	0.958	114.7	0	284 (5.80/f)

Kubric-4D & SpatialVID use full-frame PSNR/SSIM/LPIPS as the primary recon metrics (plus moving-object mask on Kubric). On SpatialVID, Ours is best on TSED (0.897) and rotation error (23.2°). On Kubric-4D summary, Ours recon is still running; its MEt3R (0.014) is already best-in-column.

5 · Average inference time — wall-clock per clip (SLURM accounting)

Honest wall-clock seconds per clip on 1×H200, from sacct ElapsedRaw summed over the COMPLETED generation array tasks divided by the clips that method produced on that benchmark (NOT mtime deltas, which parallel array writes make meaningless). Each cell: s/clip with (s/frame) in parentheses. Sorted fastest→slowest by mean s/frame — surfacing the spread from fast feed-forward (Lyra 2.0 / Matrix-Game, ~1.5–2 s/frame) to autoregressive-memory (Voyager, 13–100 s/frame) to the per-scene Shape-of-Motion 4D-GS optimization (~62 s/frame, full per-scene fit). — = no sacct generation record in the 14-day window (never fabricated).

Method	DL3DV static s/clip (s/frame)	Kubric-4D s/clip (s/frame)	NVIDIA dyn s/clip (s/frame)	SpatialVID s/clip (s/frame)	Long-horizon s/clip (s/frame)
Matrix-Game 3.0	—	—	—	—	354 (1.55)
Lyra 2.0 · Wan2.1-14B	—	—	—	—	464 (2.03)
ViewExtrapolator	131 (1.61)	—	144 (5.99)	129 (1.61)	690 (3.03)
NVS-Solver	122 (1.51)	230 (9.59)	318 (13.24)	150 (1.88)	2288 (10.04)
TrajectoryCrafter	610 (12.46)	336 (6.86)	514 (10.50)	284 (5.80)	828 (3.63)
Ours · Wan2.1-VACE-14B OURS	758 (9.36)	289 (3.57)	795 (9.81)	389 (4.81)	2732 (11.98)
ViewCrafter	264 (3.26)	226 (9.41)	505 (21.03)	299 (3.73)	843 (3.70)
I2V base · Wan2.1-14B	670 (8.28)	—	—	—	—
FlexWorld	259 (3.20)	459 (19.12)	663 (27.64)	—	1095 (4.80)
GEN3C · Cosmos-7B	498 (6.14)	503 (20.96)	779 (32.46)	455 (5.68)	1148 (5.04)
Ours-ft · VACE-14B (long-horizon FT)	—	—	—	—	3253 (14.27)
Long-term Spatial Memory	—	—	—	—	5929 (26.00)
FrameCrafter · Wan2.1-14B	2413 (29.80)	866 (36.08)	742 (30.89)	2262 (28.27)	7296 (32.00)
Voyager · HunyuanVideo	1086 (13.41)	1901 (79.22)	2395 (99.79)	1103 (13.79)	4914 (21.55)
Shape-of-Motion · 4D-GS per-scene	—	1487 (61.98)	—	—	—

Caveats (read inline):

Wall-clock SECONDS PER CLIP on 1xH200, INCLUDING one-time model-load amortized over the generation array (not a pure forward pass).
Clip LENGTHS differ per benchmark (static/kubric/nvidia/spatialvid ~81 frames; long-horizon ~150-240) -> s/frame is the cross-benchmark-fair number; the per-method output frame count is shown.
Source = SLURM sacct ElapsedRaw summed over COMPLETED generation array tasks / clips produced. Methods with no sacct record in the last 14 days (purged) show n/a -- never fabricated.
Per-scene methods (Shape-of-Motion) include FULL per-scene 4D-GS optimization, not just a forward pass.

Generated 2026-06-28T21:49:45 from SLURM sacct (code/compute_infer_time.py → infer_time.json); per-method clip counts & sacct task counts are in the JSON.

3 · Kubric-4D angle-binned difficulty curve — per-bin table

The numbers behind the headline difficulty curve at the top. Standard GCD/AnyView protocol, binned by optical-axis change from the input view. Shown per bin: full PSNR↑ / LPIPS↓, moving-object-masked mPSNR↑, and flow-warp EPE↓ (temporal). Harder bins (120°/180° back-views) = larger viewpoint change. Best per (bin×metric) bolded.

Clips per bin — 30°: n=20 · 60°: n=20 · 90°: n=20 · 120°: n=20 · 180°: n=20

Method	30° (n=20)				60° (n=20)				90° (n=20)				120° (n=20)				180° (n=20)
Method	PSNR↑	LPIPS↓	mPSNR↑	EPE↓	PSNR↑	LPIPS↓	mPSNR↑	EPE↓	PSNR↑	LPIPS↓	mPSNR↑	EPE↓	PSNR↑	LPIPS↓	mPSNR↑	EPE↓	PSNR↑	LPIPS↓	mPSNR↑	EPE↓
GEN3C · Cosmos-7B	20.08	0.323	11.60	0.611	19.83	0.337	11.79	0.593	19.04	0.355	11.73	0.657	18.95	0.366	11.46	0.632	18.66	0.360	11.83	0.591
Voyager · HunyuanVideo	16.30	0.414	9.45	0.815	16.48	0.437	9.44	2.687	15.33	0.464	13.02	2.550	13.46	0.502	14.53	1.730	12.97	0.490	13.09	1.764
FlexWorld	16.39	0.333	13.52	0.648	15.52	0.385	12.33	0.685	14.82	0.441	11.58	1.036	14.59	0.452	11.27	0.803	14.77	0.433	11.72	1.261
ViewCrafter	25.30	0.319	11.24	7.147	25.19	0.323	11.26	7.147	24.92	0.332	11.28	7.147	24.63	0.340	11.55	7.147	24.59	0.338	11.74	7.147
FrameCrafter · Wan2.1-14B	27.61	0.267	11.05	5.226	27.19	0.290	11.29	5.447	26.62	0.329	11.29	6.074	26.18	0.353	10.80	6.622	25.86	0.370	10.17	7.206
NVS-Solver	26.18	0.258	14.39	3.787	25.56	0.273	14.18	3.882	24.68	0.293	13.58	4.276	24.58	0.296	13.42	4.051	24.52	0.291	13.92	4.173
TrajectoryCrafter	—	—	—	4.465	—	—	—	5.469	—	—	—	4.971	—	—	—	4.907	—	—	—	4.481
Shape-of-Motion · 4D-GS	13.45	0.559	8.13	2.145	12.63	0.591	6.39	2.173	12.38	0.606	6.03	2.330	12.65	0.588	5.97	2.337	13.32	0.570	6.82	2.325

STANDARD GCD/AnyView angle-binned protocol; PERFECT-GT (no depth / scale / warp confound). Harder bins = larger optical-axis change from the input view (120°/180° are back-views). PSNR/LPIPS degrade monotonically with angle for warp-based methods (e.g. ViewCrafter 25.3→~18 dB), while moving-object mPSNR and flow-EPE expose temporal/dynamic fidelity. Ours (vace14b) is not yet scored on the binned recon metrics here (recon n=0) and is omitted from this table.

4 · Long-horizon geometric stability

4a · Drift / loop-closure (GVS-style, DL3DV)

Rolled-out long trajectories. ATE & loop-closure error are fractions of scene extent. Sorted by ATE (best first). FID / MEt3R folded in from the lh_dl3dv recon metrics.

Method	drift ATE↓ (frac extent)	rot drift°↓	loop-closure↓ (frac extent)	FID↓	MEt3R↓	n	infer s/clip↓
Voyager · HunyuanVideo	0.03%	0.5	0.999	—	0.059	33	4914 (21.55/f)
GEN3C · Cosmos-7B	2.40%	27.3	0.959	—	0.096	42	1148 (5.04/f)
FlexWorld	3.19%	4.4	0.925	86.9	0.093	42	1095 (4.80/f)
Ours · Wan2.1-VACE-14B OURS	3.48%	2.5	—	50.9	0.113	42	2732 (11.98/f)
ViewCrafter	6.00%	24.1	0.856	89.4	0.134	42	843 (3.70/f)
TrajectoryCrafter	12.64%	116.2	0.731	100.2	0.090	42	828 (3.63/f)
NVS-Solver	13.04%	68.0	0.407	86.4	0.109	42	2288 (10.04/f)
ViewExtrapolator	18.40%	67.5	0.568	—	0.175	42	690 (3.03/f)
FrameCrafter · Wan2.1-14B	29.78%	52.6	—	—	0.183	42	7296 (32.00/f)
Ours-ft · VACE-14B (long-horizon FT) ft@4000 · re-running	—	—	—	—	—	3	3253 (14.27/f)

Single-shot SVD methods (nvs_solver / viewextrapolator) and trajcrafter / framecrafter collapse rotationally (52–116° drift); warp-anchored VACE stays geometrically stable (3.5% ATE, 2.5° rot). voyager’s ≈0 ATE comes from near-static rollouts (almost no camera motion). The vace14b_ft_lh long-horizon-finetuned variant (ft@4000) has only .bak outputs — metrics are re-running, shown as em-dashes.

4b · WorldScore (ICCV 2025 Stanford world-gen benchmark)

Composite world-generation score. Sorted by WorldScore (best first).

Method	WorldScore↑	cam-ctrl↑	3D-consist↑	photo↑
NVS-Solver	70.44	69.43	91.63	81.23
Ours · Wan2.1-VACE-14B OURS	67.82	78.26	85.40	88.88
FlexWorld	65.08	79.48	88.26	89.60
Voyager · HunyuanVideo	63.29	74.99	80.46	83.46
ViewExtrapolator	61.15	76.64	79.52	55.97
TrajectoryCrafter	56.28	33.22	77.24	82.07
GEN3C · Cosmos-7B	44.39	53.35	53.19	38.83
FrameCrafter · Wan2.1-14B	41.14	50.95	37.22	0.00
ViewCrafter	32.27	20.91	32.36	0.00

Ours is 2nd overall (67.8) and best-but-one on camera control (78.3); nvs_solver leads overall (70.4). WorldScore aggregates camera controllability, 3D consistency and photometric quality across world-generation prompts.