README.md

---
language:
- en
pretty_name: Movie Gen Video Benchmark
dataset_info:
  features:
  - name: prompt
    dtype: string
  - name: video
    dtype: binary
  splits:
  - name: test_with_generations
    num_bytes: 16029316444
    num_examples: 1003
  - name: test
    num_bytes: 113706
    num_examples: 1003
  download_size: 16029724908
  dataset_size: 16029430150
configs:
- config_name: default
  data_files:
  - split: test_with_generations
    path: data/test_with_generations-*
  - split: test
    path: data/test-*
---

# Dataset Card for the Movie Gen Benchmark
[Movie Gen](https://ai.meta.com/research/movie-gen/) is a cast of foundation models that generates high-quality, 1080p HD videos with different aspect ratios and synchronized audio. 
Here, we introduce our evaluation benchmark "Movie Gen Bench Video Bench", as detailed in the [Movie Gen technical report](https://ai.meta.com/static-resource/movie-gen-research-paper) (Section 3.5.2).

To enable fair and easy comparison to Movie Gen for future works on these evaluation benchmarks, we additionally release the non cherry-picked generated videos from Movie Gen on Movie Gen Video Bench.

## Dataset Summary
Movie Gen Video Bench consists of 1003 prompts that cover all the different testing aspects/concepts:

1. human activity (limb and mouth motion, emotions, etc.)
2. animals
3. nature and scenery
4. physics (fluid dynamics, gravity, acceleration, collisions, explosions, etc.)
5. unusual subjects and unusual activities.
Besides a comprehensive coverage of different key testing aspects, the prompts also have a good coverage of high/medium/low motion levels at the same time.


![image/png](https://cdn-uploads.huggingface.co/production/uploads/604f82d33050a33ebb17ef65/C4Qc-4OdYRI3Oghah7fWv.png)

![image/png](https://cdn-uploads.huggingface.co/production/uploads/604f82d33050a33ebb17ef65/IJY9GUgGGRs5dDGMF2jgs.png)

## Dataset Splits
We are releasing two versions of the benchmark:
1. Test (test): This version includes only the prompts, making it easier to download and use the benchmark.
2. Test with Generations (test_with_generations): This version includes both the prompts and the Movie Gen model’s outputs, allowing for comparative evaluation against the Movie Gen model.

## Usage

```python
from datasets import load_dataset

# to download only the prompts
dataset = load_dataset("meta-ai-for-media-research/movie_gen_video_bench", split="test", streaming=True)
for example in dataset:
  print(example)
  break

# to download the prompts and movie gen generations
dataset = load_dataset("meta-ai-for-media-research/movie_gen_video_bench", split="test_with_generations", streaming=True)
for example in dataset:
  break

# to display Movie Gen generated video and the prompt on jupyter notebook
import mediapy

with open("tmp.mp4", "wb") as f:
    f.write(example["video"])

video = mediapy.read_video("tmp.mp4")
print(example["prompt"])
mediapy.show_video(video)
```

## Licensing Information
Licensed with [CC-BY-NC](https://github.com/facebookresearch/MovieGenBench/blob/main/LICENSE) License.