Dataset Viewer

The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.

StrideWorld: Benchmarking Scene Rigidity and Action Consistency in Interactive Video World Models — Data

Code · Leaderboard

The authors are Chuang Wang, Rui Nie, Jie Song, Junyuan Gao, Jiahe Song, Haitao Zhou, Boyu Zhu, Xing Li, Jiang Wu, Conghui He, Qian Yu, and Weijia Li. The arXiv link and BibTeX entry will be added after the preprint is posted.

Data release for the StrideWorld benchmark. It holds the benchmark inputs and the measurements of all 54,600 generated videos (15 models, 3,640 requests each). The companion code release rebuilds Table 2 and the fine-tuning table (Table J.1 of the appendix) from these files (bash scoring/reproduce_table2.sh <this folder>).

StrideWorld tests an interactive video world model on three properties while actions drive the camera: whether the scene keeps its 3D structure (Rigidity), whether a held action keeps the stride of its first steps (Pace), and whether the strides of single actions predict where a route ends (Composition).

Contents

start_views/
  index.json              140 start views: dataset, scene, size, share of valid depth, files with SHA-256
  views.csv               the same as a table
  SHA256SUMS              hashes of every file under cases/
  cases/<scene>/<case>/   rgb.png, depth_cm.npy (float32, cm, camera z), valid_mask.npy, camera.json (K, axes, units),
                          depth_encoding.json, provenance.json, anchors.npz
requests/requests.jsonl   3,640 requests, the same for every model
measurements/<model>/     per-video measurements of one model (15 models)
measurements/finetune/    the fine-tuning study (Sec. 6): one full benchmark run per folder
seeds/                    two further seeds of the 240 routes of the first view of every scene, all 15 models
perturbations/            controlled changes of ideal renderings with known effects (validation appendix)
results/                  table2.csv, finetune_table.csv, f2_ratio.csv, rigid_twin_ratio.csv, leaderboard.json, leaderboard.html

Download scope

For table reproduction, download only measurements/, results/, seeds/ and perturbations/ (about 300 MB). The start_views/ directory is about 6.6 GB and is needed only when running the measurement pipeline.

Start views

104 views of 24 Unreal Engine 5 scenes (1920×1080, rendered depth, exact intrinsics) and 36 photographs of 6 ETH3D scenes (full-resolution DSLR images, depth from the laser scans, valid where the scan covers the image). Every view is its own reference frame with OpenCV axes (x right, y down, z forward), centimetres, and the first pixel centre at (0, 0).

Anchors

anchors.npz in each start-view folder holds the 3D anchors that every model is evaluated with: anchor_id, uv_px (pixel position in the start view), depth_cm, xyz_cm (start-view camera frame), is_pose_anchor (pose anchors recover the camera, the others are the held-out anchors of the Rigidity test) and meta_json. Anchors are SIFT keypoints whose neighbouring depths lie within 5% of their own, or Shi-Tomasi corners on ETH3D views whose laser depth is sparse, at most 12 per cell of an 8 x 6 image grid; sorted by a hash of anchor_id, they alternate between pose and held-out anchors. A view has a median of 446 anchors (33 to 576). To reproduce the published measurements exactly, use these files instead of sampling new anchors.

Requests

Each start view has 26 requests. The 6 single actions (W forward, S backward, A left, D right, YL yaw left, YR yaw right) are held for 16 steps with seeds 0, 1 and 2. The 8 routes of 8 to 14 steps use seed 0 and form 4 pairs (G1: a route and its return, G2 to G4: swap pairs with the same actions in a different order). Fields of requests.jsonl: request_id, view, dataset, scene, seed, route_id (atomic_<action> or <pair_id>_<A|B>), family, pair_id, member, n_steps, actions. view is <scene>/<case>. The measurement files join on (scene_id, case_id, seed, action) in the CSV files and on (scene, case, seed, action or pair_id with member) in scores.json.gz. record_id identifies the generated video.

Measurements (per model)

File Rows Content
scores.json.gz 3,640 One record per video. Single actions: action, response and direction at the checkpoints, measurable12/16, per-video score. Routes: pair_id, member, steps, unit_boundary_frames, quality gates (first_frame_pass, calibration_pass, boundary_pose_pass, motion_pass, failed_motion_steps), route_valid, observed and predicted end pose (observed_endpoint_se2, predicted_endpoint_se2), errors in cm and degrees and in strides (route_translation_r, route_yaw_r), stride scales, pair errors
main_per_video.csv 2,520 Inputs of the Pace score for the single-action videos (main12, main16, aggregation weight)
unit_estimates.csv 2,520 The stride of each single-action video: rho_x_cm, rho_z_cm, theta_rad, translation_cm, yaw_abs_deg, signed_command_component
observations.csv.gz 40,320 Recovered camera after each of the 16 steps of every single-action video (t_x, t_z, theta_y_deg, measurable, frame_index)
initial_view_depths.csv 140 Median depth of the start view in the crop of the model (normaliser of the response gate), crop transform
per_checkpoint_scores.csv 5,040 Scores at steps 12 and 16 (13 models)
rigidity_perframe.csv.gz 27 k to 225 k Rigidity, per frame of the route videos: mean, median and 90th percentile of the offset of held-out anchors from their rigid positions (degrees), the recovered camera, and the image motion a rigid scene would show (par_med_deg)
rigidity_videos.csv.gz up to 1,120 Route videos with per-frame Rigidity data and their frame counts
rigid_twin_perframe.csv.gz 580 to 4,780 The same per-frame measurement on rigid renderings of the start view along the recovered camera paths of 20 route videos of the model

Model keys: matrix-game-2, matrix-game-3, matrix-game-3.5, sana-wm, dreamx-world, hunyuan-world-1.5 (HY-World 1.5), hy-world-2, lingbot-world-v2, alaya-world, h3-world, lyra2, evoke-turbo, worldcrafter-fast, solarwm-h3, hy-worldplay. Paths inside the files that pointed to our storage are reduced to <path>/<file name>.

Fine-tuning study

measurements/finetune/<run>/ has the same files for one full benchmark run each: n0 and b2 are the base and the fine-tuned model of HY-World 1.5, sana-n0 and sana-sft those of SANA-WM, hy15ref is the released HY-World 1.5 scored again in the same job, and j256 is HY-World 1.5 fine-tuned on half-step clips. The Released columns of the fine-tuning table (Table J.1 of the appendix) use measurements/hunyuan-world-1.5 and measurements/sana-wm.

Further seeds

seeds_summary.csv gives route outcomes and the Composition score per model and seed. seed_spread.csv gives, for every route completed under all three seeds, the distances between the three ends and the end error of each seed in strides. repeatability.csv, route_coverage.csv and regeneration_comparison.csv hold the agreement of route outcomes between seeds, the outcome of every route under each seed, and the comparison of the swap gap with the gap between two generations of the same route.

Controlled changes

Per-item tables of ideal renderings in which one property is changed by a known amount: m_held.csv, m_routes.csv and m_pairs.csv (pace growth, carry-over, shortened first steps, stalls), ga_videos.csv and perturbation.csv (scene deformation and appearance changes), camera_error.csv (error of the recovered camera against the true path), with their summaries (m_summary.csv, ga_summary.csv, ga_slopes.csv, perturb_ga_relative.csv, flag_rates.csv) and the comparison of generated videos with their rigid renderings under a reconstruction-based error (real_vs_twin_videos.csv, real_vs_twin_models.csv).

Licence and terms

  • Rendered UE5 start views, requests, measurements and results: CC BY 4.0.
  • ETH3D start views are derived from the ETH3D benchmark and remain under its terms. Please also cite ETH3D.
  • The generated videos are outputs of third-party models under their own licences. The full generated videos are not included in the release; the companion leaderboard includes eight annotated short clips as examples. This data repository does not contain the clips.

Citation

Chuang Wang, Rui Nie, Jie Song, Junyuan Gao, Jiahe Song, Haitao Zhou, Boyu Zhu, Xing Li, Jiang Wu, Conghui He, Qian Yu, and Weijia Li. StrideWorld: Benchmarking Scene Rigidity and Action Consistency in Interactive Video World Models. The arXiv identifier and BibTeX entry will be added after posting.

Downloads last month
45

Space using strideworld-release/strideworld 1