CineGen — Eval Pack

Val-split data used by the CineGen evaluation pipeline.

Layout

index.jsonl                  # one entry per clip (clip_id, motion_caption, logline, aspects)
matrices/<clip_id>.npz       # real 4×4 c2w GT trajectories (key: "data")
depth/<clip_id>.npy          # 128-D depth features (used by attribute classifiers)
clip_movie_mapping.json      # subset of labeled clips for attribute eval
held_out_splits.json         # per-setting held-out clip_id lists

Each index.jsonl row::

{
  "clip_id": "...",
  "dataset": "movieshots",
  "motion_caption": "camera dollies forward ...",
  "logline_script": "INT. PARK - DAY - ...",
  "macro_type": "Exterior (Open)",
  ...
}

License: CC BY-NC 4.0. See the GitHub repo for evaluation instructions.

The data is for research use only and derived from the following source datasets:

ShotBench, CineTechBench, VADB, MovieShots, CondensedMovies. See each source's terms of use for their respective licenses.