anonymous-webtoon-sfx/soundsense-webtoon
Webtoon Sound-Moment Annotations Human-curated annotations of where a vertical-scroll webtoon should make a sound, with a reference audio clip attached to each moment. Built to evaluate whether vision-language models can judge, from a static drawing alone, that a depicted moment is audible. Companion resource to the paper "SoundSense: Visual Sound Grounding in Comics and Webtoons". Webtoons have no prior sound-effect annotation, and unlike manga onomatopoeia these labels are… See the full description on the dataset page: https://huggingface.co/datasets/anonymous-webtoon-sfx/soundsense-webtoon.
Webtoon Sound-Moment Annotations
Human-curated annotations of where a vertical-scroll webtoon should make a sound, with a reference audio clip attached to each moment. Built to evaluate whether vision-language models can judge, from a static drawing alone, that a depicted moment is audible. Companion resource to the paper "SoundSense: Visual Sound Grounding in Comics and Webtoons".
Webtoons have no prior sound-effect annotation, and unlike manga onomatopoeia these labels are made from depicted events, not drawn sound words — so a moment can be marked even when the artist lettered nothing.
What is included
- `annotations.jsonl` — 359 curated sound moments. Each is a strip-space bounding box with a category, a short sound label, a trigger point (where in the box the sound lands while scrolling), whether drawn onomatopoeia happens to be present (
has_sfx_text), and one or more reference audio clips. - `audio/` — 421 reference clips (420 generated, 1 uploaded).
- `works.json` — per-work metadata and the public source URL.
Images are not included. These are coordinate-and-label annotations describing the source webtoons; use source_url to view the originals.
Works
Schema (annotations.jsonl)
{
"work": "Chailang", // romanized title
"work_original": "柴郎",
"episode": "序章-執",
"source_url": "https://mojoin.com/comics/63#chapter",
"box": {"x": 284, "y": 187, "w": 717, "h": 543}, // pixels in the stitched vertical strip
"category": "impact", // impact|motion|voice|animal|ambient|music
"sound": "sword clash", // short human label
"has_sfx_text": false, // did the artist draw a sound word here?
"sfx_text": "", // that lettering, if any
"trigger_frac": 0.5, // 0=top .. 1=bottom of the box
"note": "",
"audio": [
{"file": "audio/<hash>__name.mp3", "sha1": "...", "bytes": 12345,
"generated": true, "prompt": "a single sharp metallic clash ..."}
],
"audio_default": "audio/<hash>__name.mp3" // clip used for preview
}Boxes are in the coordinate system of the stitched vertical strip; per-episode strip dimensions and segment counts are in works.json.
Provenance and licensing
- Annotations (boxes, labels, prompts): released under CC-BY-4.0. They were curated by a professional annotator, who reviewed every strip segment in full; a VLM was used only to accelerate candidate marking, and every moment was confirmed, positioned and given a reference clip by hand.
- Reference audio: 420 clips were generated with ElevenLabs (sound-effects/v2) under a paid plan and are redistributed here; each carries the exact
promptthat produced it, so any clip is reproducible. 1 clip(s) were uploaded by the annotator. - Source images: not redistributed. The webtoons are © their respective creators and publisher (mojoin); we release only derived coordinates/labels and link to the source.
Citation
Please cite the accompanying paper (see the repository).
