VP-WorldBench

VP-WorldBench is a multi-view 4D benchmark for video world models, created as part of the Video-Point Model (VPM) project (paper · project page · code). It tests whether generated videos stay consistent across views and move like the real 3D scene.

It has 1,200 synchronized multi-view clips in four subsets: simulated and real, robot and general scenes. Every clip comes with:

• 🎥 synchronized multi-view videos (2–4 views)
• 📷 camera intrinsics and per-frame world-to-camera poses of every view
• 📍 ground-truth 3D point tracks with per-view visibility (256–1,536 points per clip)
• 📝 language captions
• 🗺️ dense depth for Simul-Robot, Simul-General and Real-General, plus subset-specific labels: object and robot labels, joint actions, person masks and 2D/3D keypoints, and quality flags

Rows from top: Simul-Robot, Simul-General, Real-Robot, Real-General. Each row shows all views of one clip at the same instant.

Subsets

subset	source	clips	views	fps	resolution	frames / clip	3D points / clip	depth
Simul-Robot	Unreal Engine, GR1T2 humanoid tabletop pick & place	500	3	20	1280×1280	81–680	768–1,536	rendered
Simul-General	CARLA, pedestrians (332) and vehicles (168) seen by following cameras	500	4	15	1280×1280	81–181	256	rendered
Real-Robot	RB-Y1 dual-arm robot with dexterous hands, teleoperated, ego + exo cameras	100	2	15	1280×720	139–325	768	–
Real-General	one person in a 47-camera capture studio, 4 selected views	100	4	15	1024×768	150	937–1,061	reconstruction

All clips are test clips.

Ground truth

subset	3D tracks	visibility	depth
Simul-Robot	exact: surface samples of the replayed objects and of the robot links	renderer depth test	rendered
Simul-General	exact: skinned surface samples of the followed agent	depth test against the rendered depth	rendered
Real-Robot	forward kinematics of the measured joints (arm + hand surface points), time-aligned to the video	robot self-occlusion	–
Real-General	pseudo-GT: surface tracks triangulated from up to 46 views (median 1.6 mm cross-validation error) + body-model tracks	z-buffer against the per-frame reconstruction	multi-view reconstruction

All 1,200 clips pass the toolkit's deep validation:

• every video frame and depth image decodes;
• every per-clip file matches its checksum;
• cameras and track invariants (visible ⊆ in-frustum ⊆ valid) hold;
• depth agrees with the tracks.

Known caveats are machine-readable in sample.json flags, with their rules in attributes.flag_rules. Examples are a target hidden in a view at the conditioning frame, or a manually fitted head-camera extrinsic on Real-Robot.

Dataset Structure

VP-WorldBench/
├── README.md
├── index.json                       # one record per clip: attributes, flags, frame counts, query counts
├── assets/example.png
└── <Subset>/                        # Simul-Robot, Simul-General, Real-Robot, Real-General
    ├── _source/                     # subset-level provenance (Real-General: caption process and accuracy)
    └── <clip_id>/
        ├── sample.json              # metadata, provenance, attributes, quality flags, file checksums
        ├── videos/view{k}.mp4       # H.264, one file per view
        ├── cameras.npz              # K (V,3,3), w2c (V,T,4,4)
        ├── tracks3d.npz             # xyz (T,N,3) metres, valid (T,N), in_frustum / visible (T,N,V), groups, parts, kinds
        ├── captions.json            # caption (+ instruction / time-segmented captions / structured attributes)
        ├── depth/depth.zip          # 16-bit PNG per view and frame (Simul-*, Real-General)
        ├── labels/                  # subset-specific labels
        └── source/                  # original per-clip files (provenance)

Conventions.

• The world frame is right-handed, in metres, with +z up.
• Cameras use OpenCV axes and world-to-camera matrices for every view and frame.
• Intrinsics refer to the stored video, with integer pixel coordinates at pixel centres.
• Depth is planar camera z.

The full specification is in docs/DATA_FORMAT.md.

size	Simul-Robot	Simul-General	Real-Robot	Real-General	total
all files	95 GB	160 GB	3.2 GB	15.6 GB	274 GB
of which depth	79 GB	106 GB	–	11.4 GB	196 GB

Usage

# everything (~274 GB)
hf download rooty2020/VP-WorldBench --repo-type dataset --local-dir data/VP-WorldBench
# or without depth (~77 GB)
hf download rooty2020/VP-WorldBench --repo-type dataset --local-dir data/VP-WorldBench --exclude "*/depth/*"
# or one subset
hf download rooty2020/VP-WorldBench --repo-type dataset --local-dir data/VP-WorldBench --include "Real-Robot/*" --include "index.json"

The toolkit (vp-worldbench, package vpm_worldbench, CLI vpmwb) reads the data, exports model inputs and scores predictions:

pip install -e "vp-worldbench[metrics,fvd]"
vpmwb validate --root data/VP-WorldBench

from vpm_worldbench import Benchmark
s = Benchmark("data/VP-WorldBench").sample("Real-Robot/episode_000")
frames = s.frames(view=0, indices=[16, 17, 18], size=(848, 480))   # RGB uint8
K, w2c = s.K(0, size=(848, 480)), s.w2c(0, [16])                   # intrinsics follow the resize
tracks = s.tracks                                                   # xyz, valid, in_frustum, visible, kind, ...
uv, vis, in_frame = s.tracks2d(0, size=(848, 480))                  # projected GT 2D tracks
depth = s.depth(0, [16], size=(848, 480))                           # metres (None for Real-Robot)

Benchmark

Multi-view video generation (protocol vpm-v1). Given frame 16 of every view, generate the next 64 frames of every view. The generated views are scored on three axes:

• appearance: PSNR, SSIM, LPIPS and FVD;
• cross-view consistency: Cross-view MEt3R. This is the distance between the generated clip's cross-view MEt3R and the real video's, averaged over clips. Each clip's MEt3R is the mean over all view pairs and timesteps.
• 3D motion accuracy: δ. The generated views are lifted to 3D tracks (MVTracker + MoGe-2) and compared with the GT tracks:
  • δ_avg counts every frame of every point;
  • δ mov. counts only the frames where the GT point has moved, so a static video scores 0.

The protocol can be changed: reference frame, I2V/V2V context length, horizon, frame step, views and resolution.

Image-to-video baselines are scored on all clips. Each view is generated independently from its own first frame. Cross-view MEt3R is lower-better; the real-video row shows its raw MEt3R.

subset	method	FVD ↓	PSNR ↑	Cross-view MEt3R ↓	δ_avg ↑	δ mov. ↑
Simul-Robot	real video / frozen points	–	–	(0.428)	60.7 / 84.5	22.0 / 0.0
	HunyuanVideo-1.5 (480p I2V, step-distilled)	98.3	20.19	0.022	45.3	4.8
	Wan2.2-TI2V-5B	314.8	15.79	0.108	32.2	2.9
	Cosmos3-Nano	104.6	20.02	0.040	43.3	6.4
Simul-General	real video / frozen points	–	–	(0.482)	60.1 / 7.3	55.7 / 0.0
	HunyuanVideo-1.5 (480p I2V, step-distilled)	115.8	11.45	0.061	25.2	17.8
	Wan2.2-TI2V-5B	206.1	10.43	0.076	19.9	12.2
	Cosmos3-Nano	167.4	10.99	0.062	24.3	17.0
Real-Robot	real video / frozen points	–	–	(0.499)	75.1 / 61.9	48.2 / 0.0
	HunyuanVideo-1.5 (480p I2V, step-distilled)	426.4	16.03	0.102	34.6	24.3
	Wan2.2-TI2V-5B	550.5	14.58	0.159	26.5	13.6
	Cosmos3-Nano	326.1	16.36	0.099	37.2	16.1
Real-General	real video / frozen points	–	–	(0.591)	68.0 / 37.6	48.0 / 0.0
	HunyuanVideo-1.5 (480p I2V, step-distilled)	94.7	20.75	0.014	29.9	3.6
	Wan2.2-TI2V-5B	122.6	17.46	0.030	28.2	4.0
	Cosmos3-Nano	56.2	21.80	0.015	31.9	3.9

Full tables, with SSIM, LPIPS, AJ, OA, survival, ADE, standard errors and the Cross-view MEt3R floors, are in the toolkit README.

Further tasks on the same data, each with input export, baselines and an evaluator in the toolkit:

task	input → prediction	metrics
3D point tracking / 3D motion	multi-view (or one) video + query points → 3D tracks	δ_avg, AJ, OA, survival, ADE, scene flow
2D point tracking	one view + 2D queries → 2D tracks; cross-view correspondence	TAP-Vid AJ, <δ^x_avg>, OA
video depth	one view → per-frame depth	AbsRel, δ<1.25, RMSE
4D reconstruction	multi-view video → cameras + point maps	RRA/RTA, ATE, RPE, Chamfer, F-score
novel-view synthesis	other views + target camera → held-out view	PSNR, SSIM, LPIPS
segmentation	video + click/box prompt → masks	J&F, point precision/recall/IoU
video understanding	video(s) + question → answer	QA accuracy, caption retrieval, temporal grounding

Sources

subset	derived from
Simul-Robot	renders of humanoid (Fourier GR1T2) pick & place episodes in Unreal Engine scenes, produced for this benchmark
Simul-General	CARLA renders of pedestrians and vehicles with target-following cameras, produced for this benchmark
Real-Robot	the RB-Y1 WujiHand2 teleoperation dataset (100 episodes, ego + exo), with 3D arm and hand point tracks from forward kinematics; videos and tracks time-aligned for this benchmark
Real-General	100 ten-second clips of an in-house 47-camera capture, with surface and body-model tracks, reconstruction depth, SAM2 person masks, 2D/3D keypoints and motion captions produced for this benchmark

The Real-General captions were written by vision-language-model annotators and checked by model verifiers. Real-General/_source/CAPTIONS.md gives the process and the measured accuracy.

People. Real-General shows real, identifiable people. Use the data only for research on video and 4D perception and generation. Do not use it to identify or profile the people shown. To request the removal of a clip, contact the maintainer: Byungwoo Jeon, imbw2024@kaist.ac.kr.

Citation

@article{vpm2026,
  title   = {Learning the 4D World through Point Trajectory Forecasting},
  author  = {Jeon, Byungwoo and Hong, Sunghwan and Kim, Dohyeon and Lee, Jonghoon and Park, Jungwoo and Lee, Inhee and
             Baek, Doosan and Kim, Jinil and Lim, Taehyun and Joo, Hanbyul and Pollefeys, Marc and Xie, Saining and
             Shin, Jinwoo},
  journal = {arXiv preprint arXiv:2610.06940},
  year    = {2026}
}

License

VP-WorldBench is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. Copyright Byungwoo Jeon. The evaluation toolkit is released separately under Apache-2.0. Simulator assets shown in the renders remain under the terms of their owners.