---
license: cc-by-nc-4.0
pretty_name: CineBenchSyn
language:
  - en
task_categories:
  - text-to-video
tags:
  - video-generation
  - cinematic-video-generation
  - multi-subject
  - temporal-control
  - benchmark
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files: metadata.jsonl
---

# CineBenchSyn

**CineBenchSyn** is the synthetic benchmark for **[CineOrchestra](https://snap-research.github.io/CineOrchestra/)**,
a unified model for cinematic video generation that jointly controls subjects, events, camera, and shot transitions.

It contains **512 hand-authored 10.2-second scenarios** that target under-represented, edge-case
cinematic compositions (large casts, dense events, frequent shot transitions). Each scenario is
expressed with the same entity-centric primitive used by CineOrchestra: every cinematic element —
a character, an object, the `{camera}`, or a `{transition}` — is an entity acting over a temporal
interval, described by `(start_time, end_time, prompt)` and, for visual subjects, a reference image.

- 📄 Paper: [CineOrchestra](https://arxiv.org/abs/2606.13768)
- 🌐 Project page: https://snap-research.github.io/CineOrchestra/

## Dataset structure

```
annotations/<id>_ultra_dense.json     # one entity-centric annotation per scenario (512 total)
reference_images/<id>_ref_image_NN_<entity>.png   # reference image per visual entity
metadata.jsonl                        # one summary row per scenario
```

Each `annotations/<id>_ultra_dense.json` has:

```json
{
  "global_entities": [
    {"name": "{rocker_jax}",
     "description": "{rocker_jax} is a lanky man in his late twenties ...",
     "ref_image_path": "00073_ref_image_00_rocker_jax.png"}
  ],
  "dense_entities": [
    {"name": "{camera}",
     "description": "{camera} holds a slow push-in on {rocker_jax} ...",
     "time_intervals": [0.0, 4.812]},
    {"name": "{shot_transition}",
     "description": "{shot_transition} shows a hard cut.",
     "time_intervals": [2.418, 2.41801]}
  ],
  "summary": "A rock concert ..."
}
```

- `global_entities` — one entry per entity: a stable brace tag (`{name}`), a static appearance
  `description`, and, for visual subjects, a `ref_image_path` (a filename in `reference_images/`).
  Special tags `{camera}`, `{transition}`/`{shot_transition}` and `{scene_*}` carry no reference image.
- `dense_entities` — timestamped events, each `(name, description, time_intervals=[start, end])` in
  seconds. Hard cuts are near-zero-width intervals `[t, t + 1e-5]`.
- `summary` — a one-line scene summary.

## Usage

```python
from huggingface_hub import snapshot_download
path = snapshot_download(repo_id="sharathgirish/CineBenchSyn", repo_type="dataset")
```

To evaluate generated videos against these annotations (subject identity, dense-caption following,
and shot-transition timing), use the benchmark evaluation code released with CineOrchestra; point its
prompts directory at `annotations/` and its reference-image directory at `reference_images/`.

## License

Released under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/) — free to use, share,
and adapt for **non-commercial** purposes with attribution. Reference images are synthetic, generated
with a public text-to-image model.

## Citation

```bibtex
@article{girish2026cineorchestra,
  title   = {CineOrchestra: Unified Entity-Centric Conditioning for Cinematic Video Generation},
  author  = {Girish, Sharath and Chen, Tsai-Shien and Dong, Zhikang and Singhal, Mukesh and Chen, Hao and Tulyakov, Sergey and Siarohin, Aliaksandr},
  journal = {arXiv preprint arXiv:2606.13768},
  year    = {2026}
}
```