README.md

metadata

license: other
license_name: mixed-source-see-below
task_categories:
  - text-to-image
  - image-to-image
tags:
  - controlnet
  - pose-estimation
  - diffusion
  - openpose
pretty_name: Pose-ControlNet Training Data (COCO + Danbooru)

Pose-ControlNet Training Data (COCO + Danbooru subset)

Paired (image, 18-point OpenPose-style pose skeleton, caption) triples for training a pose-conditioned ControlNet, built as channel-concatenated conditioning alongside the noisy latent (same mechanism as depth-conditioned ControlNets).

This is the public subset of a larger dataset — it excludes a Human-Art-derived portion (real_human / sculpture / painting) that is kept local-only per that dataset's non-commercial, no-redistribution license.

Contents

Column	Description
file_name	The source RGB image (photograph or anime illustration)
conditioning_image	The paired pose skeleton map — same resolution/alignment as file_name, 18-point body-only OpenPose-style rendering, no hands or face
text	A caption describing appearance, clothing, setting, color, lighting, and camera technique. Captions deliberately exclude all body-position/pose language — pose information is carried entirely by conditioning_image, not the text

Sources and licensing

• COCO-WholeBody (photoreal subset): images from the COCO dataset (Flickr-sourced, subject to COCO's own terms), pose annotations sourced from COCO-WholeBody, whose annotations are licensed for research and non-commercial use only (SenseTime Research) — separate from the images' own licensing.
• Danbooru (anime subset): community-sourced illustrations pulled via the public Danbooru API. Rating-filtered to general+sensitive only (questionable/explicit excluded). An additional tag-based filter was applied to exclude any content coded as depicting minors, independent of the official content rating.

Given the mixed and partially non-commercial licensing of the source annotations, this dataset is provided for research purposes; check the original source licenses before any commercial use.

Pose skeleton schema

18-point body-only OpenPose-style skeleton (COCO-17/MPII-16 remapped to a unified schema where applicable; DWPose used for Danbooru). Hands and face are deliberately excluded — hand-keypoint detection proved unreliable on stylized art, and standardizing to body-only kept the schema consistent across all sources.

Captioning methodology

Captions were generated via Gemini 3.5 Flash-Lite (Google), following a fixed prompt enforcing:

• Medium-first opening (e.g. "A photograph of...", "An anime illustration depicts...")
• Verbatim transcription of any legible text/signage in the image
• Zero pose, posture, gesture, or body-position language of any kind
• No camera framing that implies body-crop extent (e.g. "close-up", "waist up")
• 1-3 trailing style/technique tags
• 60-124 word range

Manually reviewed at multiple sample checkpoints throughout generation; a small residual rate of incidental pose-adjacent language (~1-1.5%) may remain despite the filter, tracked via automated post-hoc checks.

Content notes

• Danbooru content was rating-filtered and additionally tag-filtered for apparent-minor content before inclusion; any content flagged by the generation model as prohibited was excluded from the dataset rather than force-captioned.
• This is training data for a pose-conditioning model, not a curated art collection — quality and content vary by source.