Ziqi1018/CineScript-train

CineScript — Train Split

The train split of CineScript, the dataset introduced in Unveiling the Value of Motion for Cinematic Camera Trajectories (NeurIPS 2026). It is the data CineGEN was trained on. The matching val split is Ziqi1018/CineScript-eval; together they contain 25,785 clips.

Paper · Code · Project page · Checkpoints

Statistics

Counted on this release.

Clips	23,207
Camera poses	3,840,301
Poses per clip	mean 165.5, median 136, range 44–901
Clips linked to movie metadata	3,232, from 1,233 films

Source dataset	dataset value	Clips
CMD (Condensed Movies)	condensedmovies	18,088
CineTechBench	cinetechbench	1,800
VADB (film-derived subset)	vadb	1,559
MovieShots	movieshots	1,432
ShotBench	shotbench	328

Layout

index.jsonl                  # one entry per clip
matrices.tar.gz              # matrices/<clip_id>.npz, camera trajectories (23,207 files)
depth.tar.gz                 # depth/<clip_id>.npy, 128-D depth features (16,892 files)
clip_movie_mapping.json      # movie attributes for the metadata-linked clips
stats.json                   # the counts above

The trajectories and depth features are shipped as archives because a Hugging Face directory holds at most 10,000 files. After extraction the layout and fields are the same as in CineScript-eval, so code written for one reads the other:

tar -xzf matrices.tar.gz && tar -xzf depth.tar.gz

./scripts/download.sh --train in the GitHub repository downloads and extracts the pack in one step.

Fields

index.jsonl

Field	Content
clip_id	Clip identifier: source video ID plus the scene or shot index assigned during processing
dataset	Source dataset, see the table above
motion_caption	Motion caption describing the camera movement
logline_script	Screenplay-style logline, [INT./EXT.] [Location] - [Time] - [Action]; empty for 15 clips
macro_type, setting_class, subject_composition, genre_vibe	Additional scene tags; not used to train CineGEN

matrices/<clip_id>.npz: data is an (N, 4, 4) array of camera-to-world matrices estimated with ViPE, inds the corresponding source frame indices. The DirSpeed and Pose9D representations are computed from these matrices by the code in the GitHub repository; CineGEN uses at most 300 steps per clip during training.

clip_movie_mapping.json: for each metadata-linked clip, the movie title, release year, genres, directors, countries and IMDb ID, retrieved from Wikidata, Wikipedia and IMDb.

Split

This is the split used to train CineGEN: all clips are shuffled with seed 42, the first 10% form the val split (CineScript-eval) and the rest this train split. The two splits do not overlap.

Videos

No video frames are included. clip_id identifies the source video (for example the YouTube ID in CMD and ShotBench, or the IMDb ID in MovieShots) followed by the scene or shot index assigned during processing; segment boundaries are not part of this release. The source videos are available from the original datasets under their own terms.

License

The annotations in this repository are released under CC BY-NC 4.0. The source videos remain under the licenses of their original datasets.

Citation

@inproceedings{zhou2026unveiling,
  title     = {Unveiling the Value of Motion for Cinematic Camera Trajectories},
  author    = {Zhou, Ziqi and Yuan, Yujian and Sevilla-Lara, Laura},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}