Omni-StoryBench

Omni-StoryBench is a context-aware omnimodal story-generation benchmark. Given the current page of an illustrated children's storybook (image + narration), book-level metadata, and a structured condition describing what should happen next, a model must generate the next page across three modalities at once: its narration text, its illustration, and a spoken character utterance (with speaker attributes).

The benchmark contains 900 rigorously validated story transitions drawn from four openly licensed children's-book collections, each with ground-truth next-page references and speech metadata.

Omnimodal generation should not be evaluated as the independent production of text, images, and speech: a system may produce plausible outputs in each modality while failing to express a coherent shared event. Omni-StoryBench evaluates whether models can continue a story coherently across image, narration, and speech.

Task

Each sample is a page transition from → to within a single book.

Input (what the model is given):

• input.current_page.image — the current page illustration
• input.current_page.text — the current page narration
• input.book_metadata — book-level ("global") metadata: genre, topic, style, narrative_tense, narrative_perspective, characters
• next_page_condition — a structured description of the next page to generate:
  • characters — per-character emotion, visibility, speech_intent, action
  • scene — the next-page scene description
  • text_instruction — instruction for the next-page narration
  • ambient_sound — ambient sound of the next page

Ground truth (the reference next page):

• ground_truth.next_page.image — reference next-page illustration
• ground_truth.next_page.text — reference next-page narration
• ground_truth.speech — the spoken utterance and its metadata: utterance, speaker, emotion, speed, pitch, gender

Speech is provided as text + attributes, not audio. The spoken line (utterance) and its speaker attributes are the ground-truth reference; this release does not ship reference audio (.wav) files.

Dataset structure

Omni-StoryBench/
├── data/
│   └── omni_storybench.parquet          # machine-readable index (one row per sample)
├── samples/
│   └── <source>/<book>/<from>__<to>.json # one self-contained JSON per sample
├── metadata/
│   └── <source>/<book>.json             # clean book-level (global) metadata
├── images/
│   └── <source>/<book>/<name>.{png,jpg,jpeg}  # page illustrations
└── texts/
    └── <source>/<book>/<name>.txt       # page narration text files

• data/omni_storybench.parquet is the canonical index used by load_dataset() and the dataset viewer (900 rows). Media columns store repo-relative paths to the files under images/ and texts/.
• samples/ holds one self-contained JSON per sample, following the paper's example structure (input / next_page_condition / ground_truth). Each sample inlines its book metadata, so a single file contains everything needed for that transition.
• metadata/ holds the same book-level metadata as standalone files (one per book).
• images/ and texts/ hold the page assets referenced by both the parquet and the per-sample JSON.

Parquet columns

Column	Description
id	"<source>/<book>/<from>__<to>"
source, book	source collection and book identifier
from_key, to_key	current / next page keys (page_00N)
current_image, current_text, current_text_path	current page image path, narration text, text-file path
next_image, next_text, next_text_path	ground-truth next page image path, narration text, text-file path
genre, topic, style, narrative_tense, narrative_perspective	book-level metadata
scene, text_instruction, ambient_sound	next-page condition
speech_utterance, speech_speaker, speech_emotion, speech_speed, speech_pitch, speech_gender	ground-truth speech
condition_characters	next-page per-character conditions (JSON string)
book_characters	book-level character profiles (JSON string)
sample_path, metadata_path	repo-relative paths to the per-sample JSON and book-metadata JSON

Per-sample JSON schema (samples/…)

{
  "id": "digitallibrary/A-Cloud-of-Trash/page_004__page_005",
  "source": "digitallibrary",
  "book": "A-Cloud-of-Trash",
  "transition": {"from": "page_004", "to": "page_005"},
  "book_metadata_path": "metadata/digitallibrary/A-Cloud-of-Trash.json",
  "input": {
    "current_page": {"image": "images/…/4.jpeg", "text": "…", "text_path": "texts/…/4.txt"},
    "book_metadata": {"genre": "…", "topic": "…", "style": "…",
                      "narrative_tense": "…", "narrative_perspective": "…", "characters": [ … ]}
  },
  "next_page_condition": {
    "characters": [{"name": "…", "emotion": "…", "visibility": "…", "speech_intent": "…", "action": "…"}],
    "scene": "…", "text_instruction": "…", "ambient_sound": "…"
  },
  "ground_truth": {
    "next_page": {"image": "images/…/5.jpeg", "text": "…", "text_path": "texts/…/5.txt"},
    "speech": {"utterance": "…", "speaker": "…", "emotion": "…", "speed": "…", "pitch": "…", "gender": "…"}
  }
}

Usage

from datasets import load_dataset

ds = load_dataset("snu-aidas/Omni-StoryBench", split="test")
ex = ds[0]
print(ex["id"], ex["text_instruction"])
print(ex["current_image"], "->", ex["next_image"])

Image and text columns are repo-relative paths. To download a referenced asset:

from huggingface_hub import hf_hub_download

img_path = hf_hub_download(
    repo_id="snu-aidas/Omni-StoryBench",
    filename=ex["current_image"],           # e.g. "images/digitallibrary/A-Cloud-of-Trash/4.jpeg"
    repo_type="dataset",
)

Alternatively, load the fully self-contained per-sample record:

import json
from huggingface_hub import hf_hub_download

sample = json.load(open(hf_hub_download(
    "snu-aidas/Omni-StoryBench", ex["sample_path"], repo_type="dataset")))

Sources & Licensing

Omni-StoryBench is constructed from four collections of publicly available, openly licensed children's storybooks. Licenses vary by source and, for some sources, by book. Users must comply with the license and terms of use of each original source; each sample retains its source and book so provenance can be traced.

Source (source)	License	Link	# Samples
African Storybook (africanstorybook)	CC BY 4.0	https://www.africanstorybook.org/	54
Global Digital Library (digitallibrary)	CC BY / CC BY-SA (per book)	https://digitallibrary.io/	293
Storybooks Canada (storybookscanda)	CC BY / CC BY-NC (per book)	https://www.storybookscanada.ca/downloads/	7
Storyweaver (storyweaver)	CC BY 4.0	https://storyweaver.org.in/en/stories?language=English	546
Total			900

Because some source material is released under CC BY-NC (Storybooks Canada) and CC BY-SA (some Global Digital Library books), the aggregate benchmark is not uniformly permissive: treat it as a mixed-license collection and respect the most restrictive applicable license for each item. The license field is therefore set to other.

Intended use

Omni-StoryBench is intended for evaluating context-aware omnimodal generation models that jointly produce the next story page's narration, illustration, and spoken utterance, conditioned on the current page and structured next-page conditions. It is a test/evaluation benchmark (single test split of 900 samples), not a training corpus.

Limitations

• The dataset is built from children's storybooks and may not represent all cultures, narrative styles, visual styles, or speech patterns.
• Book-level metadata and next-page conditions are model-assisted annotations that were subsequently human-validated; residual noise may remain.
• Speech is represented as a spoken-line transcript plus attributes (speaker, emotion, speed, pitch, gender), not as reference audio.

Citation

The dataset accompanies the paper "What Comes Next? Omni-StoryBench for Evaluating Story-Grounded Omnimodal Generation."

@misc{hyeon2026omnistorybench,
      title={What Comes Next? Omni-StoryBench for Evaluating Story-Grounded Omnimodal Generation}, 
      author={Sieun Hyeon and Yejoon Lee and Mintaek Lim and Woojin Kim and Jaeik Kim and Jaeyoung Do},
      year={2026},
      eprint={2609.37317},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.37317}, 
}