README.md

---
license: apache-2.0
task_categories:
- image-to-text
- visual-question-answering
- text-to-image
language:
- en
tags:
- image-to-text reasoning
- sequential
- human-annotated
- multimodal
- vision-language
- movie
- scenes
- video
- frames
pretty_name: StoryFrames
size_categories:
- 1K<n<10K
---

# The StoryFrames Dataset
[StoryFrames](https://arxiv.org/abs/2502.19409) is a human-annotated dataset created to enhance a model's capability of understanding and reasoning over sequences of images.
It is specifically designed for tasks like generating a description for the next scene in a story based on previous visual and textual information.
The dataset repurposes the [StoryBench dataset](https://arxiv.org/abs/2308.11606), a video dataset originally designed to predict future frames of a video.
StoryFrames subsamples frames from those videos and pairs them with annotations for the task of _next-description prediction_.
Each "story" is a sample of the dataset and can vary in length and complexity.

The dataset contains 8,881 samples, divided into train and validation splits.
![StoryFrames dataset distribution](storyframes-scene-dist-tabblue-taborange.png)

If you want to work with a specific context length (i.e., number of scenes per story), you can filter the dataset as follows:

```python
from datasets import load_dataset

ds = load_dataset("ingoziegler/StoryFrames")

# to work with stories containing 3 scenes
ds_3 = ds.filter(lambda sample: sample["num_scenes"] == 3)
```

## What Is a Story in StoryFrames?
* **A story is a sequence of scenes:**
Each story is composed of multiple scenes, where each scene is a part of the overall narrative.

* **Scenes consist of two main components:**
  * **Images**: Each scene is made up of several frames (images) that have been subsampled from the original video.
  * **Scene Description**: There is a single textual description for each scene (i.e., one or more images) that captures the plot of the scene.

## How Is the Data Organized?
* **Temporal Markers:**
  * `start_times` and `end_times`: These fields provide the time markers indicating when each scene begins and ends in the video. They define the boundaries of each scene.

* **Frame Subsampling:**
  * `subsampled_frames_per_scene`: For each scene, a list of frame timestamps is provided. Each timestamp is formatted to show the second and millisecond (for example, `frame_sec.millisec` would be `frame_1.448629`). These timestamps indicate which frames were selected from the scene.

* **Image Data:**
  * `scenes`: In a structure that mirrors the subsampled timestamps, this field contains the actual images that were extracted. The images are organized as a list of lists: each inner list corresponds to one scene and contains the images in the order they were sampled.

* **Narrative Descriptions:**
  * `sentence_parts`: This field contains a list of strings. Each string provides a description for one scene in the story. Even though a scene is made up of multiple images, the corresponding description captures the plot progression over all images of that scene.
 
## Detailed Field Descriptions
* `sentence_parts`
  * Type: `List[str]`
  * A narrative breakdown where each entry describes one scene.

* `start_times`
  * `List[float]`
  * A list of timestamps marking the beginning of each scene.

* `end_times`
  * Type: `List[float]`
  * A list of timestamps marking the end of each scene.

* `background_description`
  * Type: `str`
  * A brief summary of the overall setting or background of the story.

* `video_name`
  * Type: `str`
  * The identifier or name of the source video.
  * This is not a unique identifier for stories as a video can contain multiple stories that are annotated separately.

* `question_info`
  * Type: `str`
  * Additional information used together with the video name to uniquely identify each story.

* `story_id`
  * Type: `str`
  * Automatically generated by combining `video_name` and `question_info` (e.g., "video_name---question_info") to create a unique identifier for each story.

* `num_actors_in_video`
  *  Type: `int`
  * The number of actors present in the video.

* `subsampled_frames_per_scene`
  * Type: `List[List[float]]`
  * Each inner list contains the timestamps (formatted as `frame_sec.millisec`, e.g., `frame_1.448629`) for the frames that were selected from a scene.
  * Each position of the inner lists correspond to the position of the description in `sentence_parts` and `scenes`,
  * The number of inner lists corresponds to the number of available `scenes`, as marked in `num_scenes`.

* `scenes`
  * Type: `List[List[Image]]`
  * Each inner list holds the actual frames (images) that were subsampled from a scene.
  * The structure of this field directly corresponds to that of `subsampled_frames_per_scene`.
  * Each position of the inner lists correspond to the position of the description in `sentence_parts` and `subsampled_frames_per_scene`.

* `num_scenes`
  * Type: `int`
  * The total number of scenes in the story.

* `caption`
  * Type: `str`
  * An optional caption for the sample.
  * This may be empty if no caption was provided.

* `sentence_parts_nocontext`
  * Type: `List[str]`
  * A variant of the scene descriptions that excludes sequential context.
  * This may be empty if no annotation was provided.

## Citation
The dataset was introduced as part of the following paper:

[ImageChain: Advancing Sequential Image-to-Text Reasoning in Multimodal Large Language Models](https://arxiv.org/abs/2502.19409)

If you use it in your research or applications, please cite the following paper:

```
@misc{villegas2025imagechainadvancingsequentialimagetotext,
      title={ImageChain: Advancing Sequential Image-to-Text Reasoning in Multimodal Large Language Models}, 
      author={Danae Sánchez Villegas and Ingo Ziegler and Desmond Elliott},
      year={2025},
      eprint={2502.19409},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2502.19409}, 
}
```