WanderBench: data, model outputs and scores

WanderBench evaluates video-input, camera-controlled world models. A model receives a recorded video and a camera trajectory driven by keyboard-style actions: W/A/S/D move, I/J/K/L look. It must re-render the same scene from the commanded camera, synchronized with the source video. Code, metric definitions and the leaderboard are on GitHub: YuhangChen1/WanderBench.

This repository has the same folder layout as the code repository. Download it into the root of a clone and every tool finds its files:

git clone https://github.com/YuhangChen1/WanderBench.git && cd WanderBench
huggingface-cli download FrancisChen1/WanderBench --repo-type dataset --local-dir .
# or from ModelScope
modelscope download --dataset FrancisChen/WanderBench --local_dir .

Contents

Path	Content
data/processed_16fps/	source clips: 16 fps masters (<clip_id>.mp4) plus PTS maps (<clip_id>.json)
data/model_inputs/<model>/<clip>/<window>/	each model's input window: input.mp4, input.json and the valid-content mask
data/catalog*.{csv,json}, data/selected_*	clip catalog: source, category, tier, duration, title, origin
outputs_v2/<model>/<clip>/<window>/<trajectory>/<seed>/	random-key outputs: output.mp4, output_metadata.json (source time of every frame), run_config.json, camera poses, prompt
harddemo/outputs/<model>/<clip>/...	long-revisit (out-and-back) outputs, same files
outputs_syncwarp/<variant>/...	WanderLand outputs, plus TrajectoryCrafter control reruns
trajectories/, harddemo/trajectories/	canonical trajectories: poses.npy, intrinsics.npy, action_sequence.json, caption
bench/jobs_manifest.jsonl	one row per job: clip, window, trajectory, model, output folder
captions/	scene captions
wbmetrics/runs/	raw per-output metric results
wbmetrics/cache.tar.gz	reference ceilings and cached VLM answers; unpack with tar -xzf wbmetrics/cache.tar.gz -C wbmetrics
reports/wbmetrics_scores_per_output.jsonl	final scores of every output: job_id, set, model, clip_id, window, output, scores{M01..M24}
reports/wbmetrics_by_difficulty.*, reports/wbmetrics_by_model_difficulty_dimension.csv	tables by model × difficulty × dimension
reports/syncwarp/	WanderLand job lists, paired results and before/after videos

• Models
  • Full length: World-in-World, InSpatio-World 1.5, WorldForge v2.
  • Frame-limited: UniWorld-View, WorldForge v1, ReCamMaster, CameraAnything (81 frames each); TrajectoryCrafter (49 frames).
• Random-key set: 257 clips, in three difficulty tiers by duration (Simple ≤ 8 s, Medium 8–15 s, Hard > 15 s). Full-length models run on every tier. Frame-limited models run on Simple.
• Long-revisit set: out-and-back trajectories on Medium and Hard clips. The camera goes out, pauses, then retraces its keys back to the start.
• Scores are on a 0–100 scale in 24 metrics and 6 dimensions. See docs/WanderBench_metrics.md in the code repository.

Sources and terms

• Real clips (dance, film, nature) are short excerpts of publicly available videos. data/catalog_full.json gives each clip's title and origin.
• Generated clips were made with MiniMax.
• Rights in the underlying footage remain with the original owners. The source clips and model outputs are provided for research and benchmarking of camera-controlled video models only.
• To have an item removed, open a discussion on this repository.

中文说明

WanderBench 评测"视频输入、相机可控"的世界模型：给定一段视频和一条由按键（WASD 移动、IJKL 转视角）驱动的相机轨迹，模型要从 新视角、与源视频同步地重新渲染同一场景。

本仓库的目录结构与 GitHub 代码仓库一致，下载到代码仓库根目录即可直接使用。内容包括：

• 源视频（16 fps）与各模型的输入窗口；
• 8 个模型的全部输出（随机按键集、长回访集）；
• WanderLand 输出；
• 轨迹、任务清单和字幕；
• 原始指标结果，以及每个输出的最终分数（reports/wbmetrics_scores_per_output.jsonl）。

真实视频片段来自公开视频，版权归原作者，仅供研究与评测使用；如需删除，请在本仓库发起讨论。