Unveiling the Value of Motion for Cinematic Camera Trajectories
Paper • 2609.38683 • Published
CineGEN — Checkpoints
Checkpoints for Unveiling the Value of Motion for Cinematic Camera Trajectories (NeurIPS 2026).
Paper · Code · Project page · Data
| Path | Description | Size |
|---|---|---|
cinegen/best.pt |
CineGEN generator: masked autoregressive model on DirSpeed tokens, conditioned on the motion caption, the logline and the first camera pose | 477 MB |
align_dirspd_motion/best.pt |
Purely contrastive alignment evaluator (DirSpeed), used for F1, FCD, Coverage, AlignScore, R@K and MedR | 514 MB |
clatr_dirspd_motion/best.pt |
CLaTr evaluator (DirSpeed), used for the independent CLaTr evaluation | 194 MB |
clf/<setting>/{direction_speed,trajectory}_best.pt |
Movie-attribute probes, 15 settings, one per representation (direction_speed = DirSpeed, trajectory = Pose9D) |
178 MB total |
The Era, Genre and Director columns of the main table use era_three_cine, genre_3_drop_ml_traj_only
and director_3_drop_traj_only. The other settings are read by evaluate/eval_attribute.py for the
extended attribute analysis.
License: CC BY-NC 4.0. See the GitHub repo for usage instructions.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support