README.md

metadata

license: other
license_name: mixed-see-below
pretty_name: KING2 Stickman Dataset
language:
  - en
tags:
  - stick-figure
  - pose
  - stickman
  - line-drawing
  - text-to-image
  - lora
task_categories:
  - text-to-image
size_categories:
  - 10K<n<100K

KING2 Stickman Dataset — 10,085 images

Curated dataset of stick-figure images with English captions, built for training the KING2-IMAGE SDXL LoRA — model RASHID778/king2-image, adapter stickman.

Two parts with different licenses — see the License section before reuse:

Range	Count	Source	License
images/000000.png – 009999.png	10,000	HumanML3D motion-capture renders (synthetic skeleton)	Research use — see below
images_commons/010000.png – 010084.png	85	Real illustrations, Wikimedia Commons	Per-image CC0/PD/CC BY/CC BY-SA — see ATTRIBUTION.md

Note: the Commons addition lives in a separate images_commons/ / captions_commons/ directory pair, not inside images/ — Hugging Face caps git directories at 10,000 files, and images/ is already exactly at that cap.

Sources

Source	Contribution
Deepthought42424/humanml3d_stick_figures_5_frames	29,228 motion sequences (832,864 frames scanned) — primary image source
hassanjbara/humanml3d_stick_figures	21 sequences (subset of the same corpus, deduplicated by sample_id)
HumanML3D texts	official motion descriptions used as captions (keyed by sample_id)
Wikimedia Commons	85 real hand-drawn / vector stickman illustrations, individually license-verified (see ATTRIBUTION.md)

A wider search of Hugging Face, Kaggle and GitHub (keywords: stickman, stick figure, line drawing, skeleton drawing, stickman pose/action…) found no other public dataset of comparable scale and usable licensing for the primary 10,000-image corpus.

Pipeline (images 000000–009999)

1. Frame selection — for every motion sequence, 3 candidate frames at 30/50/70% of the motion were decoded and the highest-scoring mid-action pose kept (1 image per sequence, maximizing pose diversity).
2. Quality scoring (1–10) — heuristic on ink ratio, contrast (std), figure bounding-box coverage and edge clipping. Only images with score ≥ 8 were kept.
3. Deduplication — exact + near-duplicate removal via 256-bit dHash with 32-bit-band LSH (Hamming ≤ 12): 5,936 near-duplicates removed (23,292 unique remained).
4. Standardization — every image converted to 1024×1024, RGB, PNG (source frames are square 512×512 renders; LANCZOS upscale — lossless in practice for line art).
5. Captioning — the official HumanML3D text for each motion, rewritten to stickman phrasing (e.g. "a person is taking a picture" → "a stickman is taking a picture, simple black and white stick figure line drawing"), plus derived action tags.
6. Final selection — top 10,000 by quality score.

Addition (images 010000–010084): real Wikimedia Commons illustrations

The pipeline above produces motion-capture-derived skeleton poses only (see Known limitations below). To add real hand-drawn/vector stickman art, Wikimedia Commons was searched (category Stick figures + full-text search for "stickman"/"stick figure"/"Strichfigur"), and every candidate's imageinfo license metadata was fetched via the Commons API and checked programmatically — only files tagged Public Domain, CC0, CC BY, or CC BY-SA (with no "trademarked"/other restriction flag) were kept. Results were then manually filtered for actual relevance (excluding lookalike matches like ancient pottery depicting a "youth with a stick", museum walking-stick artifacts, and real-world hazard-sign photography) and for safe subject matter. Each kept image was normalized to 1024×1024 RGB PNG (aspect-preserving resize, centered on a white canvas, matching the rest of the dataset) and captioned from its Commons title.

Full per-image attribution (author, license, source link) is in ATTRIBUTION.md — required reading before reuse, since CC BY / CC BY-SA require attribution and ShareAlike compliance on redistribution.

Structure

dataset/
├── images/              10,000 × 1024×1024 RGB PNG (000000.png … 009999.png) — HumanML3D
├── captions/             one .txt caption per image (same basename)
├── images_commons/        85 × 1024×1024 RGB PNG (010000.png … 010084.png) — Wikimedia Commons
├── captions_commons/       one .txt caption per image (same basename)
├── metadata.jsonl          file_name, caption, tags, source, + per-source fields (all 10,085 rows)
├── ATTRIBUTION.md           per-image author/license/source for images_commons/*
└── README.md

metadata.jsonl examples

HumanML3D-derived (000000–009999):

{"file_name": "images/000000.png", "caption": "a stickman walking in a strong manner, simple black and white stick figure line drawing", "tags": ["stickman", "stick figure", "pose", "line drawing", "minimal", "black and white", "walking"], "quality_score": 10, "source": "humanml3d_stick_figures_5_frames", "sample_id": "...", "frame_index": 12}

Wikimedia Commons addition (010000–010084):

{"file_name": "images_commons/010000.png", "caption": "a stickman illustration, simple black and white stick figure line drawing", "tags": ["stickman", "stick figure", "wikimedia-commons", "real-illustration"], "source": "wikimedia_commons", "commons_title": "File:HD@DH.nrw Strichfigur 1.svg", "commons_url": "https://commons.wikimedia.org/wiki/File:HD@DH.nrw_Strichfigur_1.svg", "license": "CC BY 4.0", "license_url": "...", "artist": "HD@DH.nrw | Michelle Dahlmanns"}

Known limitations

• Single visual style (000000–009999): all HumanML3D-derived images are thick-stroke skeleton renders on white — no head circle, no colors, no backgrounds, no objects. The 85-image Commons addition partially addresses this with real illustrated stick-figure art, but is still a small sample relative to the main corpus.
• Single figure per image (000000–009999): HumanML3D motions are single-person; no multi-figure scenes (a few images in the Commons addition do show multiple figures).
• License is mixed — read before reuse:
  • images/000000.png–009999.png: HumanML3D derives from AMASS/HumanAct12, which carry academic/research licenses. The upstream HF datasets declare no explicit license. Treat this range as research use; review licensing before commercial deployment.
  • images/010000.png–010084.png: each image is individually Public Domain, CC0, CC BY, or CC BY-SA — see ATTRIBUTION.md for the exact license and required attribution per file. CC BY-SA files impose ShareAlike obligations on redistribution of derivative datasets (not on model weights trained from them).