EthnicErotic/pose-catalog

Erotic Pose Catalog

A structured, machine-readable catalog of named erotic poses and sex positions — solo and multi-person — as they appear in adult video and print. Each position carries structural attributes, apparent participant composition, presentation metadata, cross-references, and an idealised COCO-17 keypoint template per participant.

Adults only (18+). This dataset documents adult sexual positions for research, content-tagging, and creative-tooling use. It contains no images and no personal data — only text metadata and schematic keypoint coordinates. It is tagged not-for-all-audiences.

Why this exists

Named-position knowledge and pose geometry have never lived in the same place:

• Human-pose datasets (COCO Keypoints, MPII, PoseTrack) give joint coordinates but no named positions, and none cover intimate poses.
• Named-position references (Wikipedia's "List of sex positions", Wikidata's sex position Q8394) give names and prose but no geometry and only coarse structure.
• Adult ML datasets (NPDI, LSPD) are binary or anatomy-level, with no position labels.

This catalog is the bridge: a clean named-position → structured-attributes → COCO-17 keypoint template mapping, decoupled from any single tag folksonomy.

As of v2 it is also the join table. There is no shared identifier for a named sexual position: the folksonomy that open adult checkpoints were trained on organises by act rather than by position, structured knowledge bases cover only a minority of positions, and every commercial generator ships a private dropdown with no exported ids. The crosswalk config maps each position into those vocabularies and, wherever the mapping is imprecise, states in writing what it discards.

• Reproducibility code (Apache 2.0): https://github.com/Agaveis/pose-catalog-pipeline
• Methodology paper: https://github.com/Agaveis/pose-catalog-pipeline/blob/main/paper/methodology-v1.md
• Archived pipeline (DOI): https://doi.org/10.5281/zenodo.21439174 (concept DOI, resolves to the latest version) · v2.0.0: https://doi.org/10.5281/zenodo.21868961
• Live catalog: https://ethnicerotic.com/poses

What's in v2.1

The position layer grows from 104 to 219 named positions (115 added 2026-10-12, the long tail of the industry and historical vocabularies), and every v2 table is regenerated from the same source files. Schemas are unchanged: v2.1 is additive.

• 219 named positions and 493 participant keypoint templates (COCO-17, composed from the same 17 base skeletons).
• Crosswalk: 339 rows. 219 folksonomy rows (18 exact, 7 close, 189 broader, 5 with no tag at all), 85 historical-name rows (82 close, 3 related) and 35 knowledge-base entities (31 exact, 3 broader, 1 close). The largest collapse now puts 26 catalog positions on one folksonomy tag. Two emitted tags are undocumented (live, with an empty wiki body) and their rows are held below exact.
• 165 of 219 positions have no precise external identifier: no exact or close folksonomy tag and no knowledge-base entity. For those, this catalog is the only precise record.
• Provenance: 88 links across 55 positions and 17 sources. The works index (87 artworks, 14 depicts links) is unchanged.
• action_beat_join covers all 219 positions. Solo render bindings stay at 33 rows for the 11 solo positions; the other 208 positions are deliberately unbound, for the reason given under v2.0.

What's in v2.0

Two new layers and three new binding tables. The position layer is unchanged: v2 is additive, and every v1.4 config keeps its schema.

The action layer. A named position is static, but generated adult media moved from stills to image-to-video, and in that pipeline the position is already fixed by the conditioning frame. What is left to choose is the motion. action-dynamics adds 22 action beats, 6 tempi, 14 camera moves, 4 loop types and a contact-continuity axis. Every beat and camera move carries a mechanical gloss: a clinical, physical description of what moves, in which direction, at what rate, written to be pasted into a video prompt. Video models respond to that and degrade on literary or euphemistic phrasing. The action_beat_join config records which beats are physically possible in each of the 104 positions, which one is characteristic, and the camera moves that suit the geometry.

Three context axes. Wardrobe (coverage, disarray, garment class), scene (location, lighting, production mode, time of day) and expression (gaze, affect, mouth). These are the controls every commercial generator exposes first and no structured position taxonomy carries. Coverage is deliberately separated from disarray, and gaze from affect from mouth state, because folksonomy tags bundle them and a bundled value cannot be recombined.

The crosswalk (194 mappings). 104 folksonomy rows, 31 knowledge-base entities, 59 attested historical names. Relations follow SKOS: 17 exact, 85 broader, 1 close, 1 with no tag at all. Every folksonomy tag was checked against the live tag API for existence, category and deprecation, and the decisive tag documentation was read before the mapping was accepted.

Context tag bindings. 79 of 88 context values bind to a live tag; 9 are honestly marked as having none, because a folksonomy built for anime image tagging does not name everything this vocabulary does.

Solo render bindings (33). Prompt bindings for the 11 solo positions across three model families. The other 93 positions are deliberately unbound: global pose conditioning does not hold figure count on multi-participant scenes, and shipping a guess would be worse than shipping nothing.

What v2 is honest about

• Most mappings are lossy, and say so. 85 of 104 folksonomy mappings are broader. Each carries a concrete statement of the discarded feature and a count of sibling positions sharing the same tag. The largest collapse puts 6 catalog positions on one tag.
• 61 of 104 positions have no precise external identifier anywhere: no exact or close folksonomy tag and no knowledge-base entity. For those, this catalog is the only precise record.
• Wiki sense-checks are documented for 46 of the 104 folksonomy rows, not all of them. This should not be described as verified end to end.
• Four emitted tags are undocumented (live, not deprecated, but with an empty wiki body). Rows using them are held below exact and no claim is made about their extension.
• Policy flags travel with the value. Values some platforms restrict are enumerated and flagged rather than omitted, so a consumer can exclude them by rule instead of by absence.

What's in v1.4

• NEW works config — a text-only index of 87 historically attested erotic artworks across 13 traditions, from Old Babylonian plaques and Attic kylikes through Moche vessels, Khajuraho friezes, Qajar albums, shunga, and early erotic photography. Each row carries the work's title, date, holding institution (with accession numbers where public), medium, a position-relevant description, a significance reading, and a verification confidence grade (high = independently confirmed against a primary source, 49 works; medium = core facts confirmed, 32; low = key facts unconfirmed, 6). Corrections found during fact-checking — misattributions, misdatings, positions misdescribed in secondary literature — are preserved in the note field rather than smoothed over.
• depicts links 14 works to catalog positions only where the work's own catalog description names the arrangement; every link was adversarially re-verified.
• Registry v2.1: sources gain iso_countries; five further art traditions registered (Etruscan tomb painting, West Mexican shaft-tomb ceramics, 19th–20th c. European erotica, Oceanic carving, African figure sculpture) — the last two documented as honest negatives: no confirmed two-person position depiction exists in those corpora.
• No images are stored or distributed — titles, dates, holdings, and accession numbers are facts of the public record; all prose is original.

What's in v1.3

• The provenance layer goes visual. Ten historical erotic-art traditions join the seven manuals: Old Babylonian plaques (Mesopotamia, c. 1800 BCE), the Turin Erotic Papyrus (Egypt), Greek symposium pottery (with comedy's kubda and keles), Pompeian frescoes (mulier equitans, Venus pendula), Moche erotic ceramics (pre-Columbian Peru), Khajuraho & Konark temple sculpture, Persian & Mughal miniatures, Chinese spring-palace albums, shunga woodblock prints, and I Modi. Sources now carry a type field (manual | art).
• 59 provenance links across 26 positions and 17 sources, spanning six world regions — Mesopotamia and the pre-Columbian Andes enter the record.

What's in v1.2

• 104 named positions (up from 85): the industry long tail lands — pretzel, butter churner, spider, facefuck/irrumatio, footjob, lap dance, jockey, sixty-eight, plus a full multi-partner wave (reverse gangbang, bukkake circle, double blowjob, side-by-side doggy, doggy-over-69, spoon DP, reverse DP, lying spit roast).
• NEW provenance config — historical & cultural provenance. Positions attested in historical sex manuals and erotic literature are tagged with their historical names and source cultures: the Kama Sutra and Ananga Ranga (India), the Su Nü Jing Nine Ways and Dongxuanzi (China, via the Ishimpō), the Edo-period Shijūhatte (Japan), al-Nafzawi's Perfumed Garden (Maghreb), and Ovid's Ars Amatoria (Rome). One row per (position × source), with the source registry in data/historical_sources.jsonl. All works are long in the public domain; summaries and mappings are original editorial.
• poses gains a historical_sources column (JSON array of {source, name, note}).

What's in v1.1

• 85 named positions (up from the 13-position v1 seed): full coverage of the solo-display, missionary, rear-entry, on-top, side-lying, seated, standing/carry, oral, manual/non-penetrative, and group families, including Kama Sutra classics and internet-era named positions.
• 172 participant keypoint templates — all base skeletons redrawn in profile view (v1.2.0, 17 postures) so facing and stacking are legible; per-pose articulation via scene-space joint_overrides.
• Aliases enriched from Wikidata (CC0) alternate labels plus industry/booru vernacular.

What's in v1

13-position seed batch establishing the schema across solo, one-on-one (penetrative/oral), and group positions.

Config	Rows	Description
poses	219	One row per named position: structure, composition, presentation, aliases, Wikidata anchor, historical sources, description, canonical URL
keypoints	493	One row per (pose × participant): base posture, placement transform, and the composed COCO-17 keypoints
pose_dimensions	32	Flat controlled vocabulary: every dimension across 8 facets (249 defined values), with citations
base_skeletons	17	The canonical base-posture skeletons the keypoints are composed from (data/base_skeletons.jsonl)
provenance	88	One row per (position × historical source): historical name, source title/author, culture, era, language, region — 22 sources (7 manuals + 15 art traditions)
works	87	One row per attested historical artwork: title, date, holding + accession, medium, description, significance, verification confidence, correction notes, and depicts links to catalog positions
crosswalk	339	One row per (position × namespace): foreign identifier, SKOS relation, what the mapping discards, sibling count, structural support tags
action_beat_join	219	One row per position: characteristic beat, all physically possible beats, contact continuity, suited camera moves
context_tag_bindings	88	One row per context value: the folksonomy tag that denotes it, or an explicit none
render_bindings_solo	33	11 solo positions × 3 model families: positive prompt, negative hints, expected figure count
Extra	n/a	data/vocabularies.jsonl (8 nested vocab schemas) · data/historical_sources.jsonl (source registry) · data/binding_schemas.jsonl (crosswalk + render-binding schemas)

Source & licensing

• Named positions are anchored to Wikidata items (structured data, CC0) for stable QIDs and cross-language names, supplemented by public-domain historical names (e.g. Kama Sutra) and the structural axes used across position references.
• Synonyms (aliases) draw on adult-industry and booru-style tag folksonomies.
• All descriptions are original. Taxonomy structure is referenced but share-alike (CC BY-SA) article prose is never copied, so this dataset is released under a permissive attribution licence.
• Keypoints are idealised schematic templates — not motion-captured or image-derived.

License: CC BY 4.0 (data). Wikidata anchors: CC0.

Reproducibility

The entire construction pipeline is open source under Apache 2.0 at Agaveis/pose-catalog-pipeline: the canonical vocabularies, the deterministic COCO-17 keypoint compositor, and the export scripts that produce every file in data/. It needs no database and no API keys — the same JSON that seeds the live catalog builds this dataset, so the two never diverge. Construction, schema, coverage, and limitations are written up in the methodology paper.

Configurations

poses

id, slug, name, aliases, wikidata_qid, kama_sutra_name, position_family, primary_act, arrangement, penetration_type, facing, dynamics, support_surface, participant_count, composition_code, n_participants, source_medium, camera_framing, difficulty, prevalence, intensity_tier, description, wiki_url, canonical_url

keypoints

pose_id, pose_slug, pose_name, participant_index, role, base_posture, placement, keypoints_coco

• keypoints_coco — flat array of 51 numbers [x, y, v, ...] in COCO-17 joint order, normalised to [0,1], origin top-left. v: 0 = unlabelled, 1 = occluded, 2 = visible.
• placement — the transform (translate, scale, rotate_deg, flip_x) applied to the base_posture skeleton to produce the scene keypoints.
• The JSONL variant additionally provides a named joints object.

pose_dimensions

facet, facet_display, dimension_id, dimension_name, dimension_type, scale, scale_citation, description, value_count, value_ids, license, version — the controlled vocabulary that every attribute value in poses comes from.

base_skeletons

posture, joints — the 17 canonical COCO-17 skeletons (supine, prone, kneeling, all_fours, astride, inverted, …) that participant keypoints are composed from.

provenance

pose_id, pose_slug, pose_name, source_id, source_title, source_author, source_culture, source_era, source_language, source_region, historical_name, note, pose_canonical_url — the historical/cultural provenance layer. Positions attested in historical sex manuals carry their historical names (e.g. Dhenuka, Clinging Cicada, Matsuba-kuzushi, the Stopperage), tying the modern vocabulary to the Indian, Chinese, Japanese, Maghrebi, and Roman literary traditions that first recorded it. Useful for cross-cultural studies of erotic vocabulary and for culturally-grounded content curation.

Quick start

from datasets import load_dataset

poses = load_dataset("EthnicErotic/pose-catalog", "poses")["train"]
kps   = load_dataset("EthnicErotic/pose-catalog", "keypoints")["train"]

row = poses[0]
print(row["name"], row["composition_code"], row["arrangement"])

import json
kp = kps[0]
coco = json.loads(kp["keypoints_coco"]) if isinstance(kp["keypoints_coco"], str) else kp["keypoints_coco"]
assert len(coco) == 51  # 17 joints × (x, y, v)

The COCO-17 joint order is: nose, left_eye, right_eye, left_ear, right_ear, left_shoulder, right_shoulder, left_elbow, right_elbow, left_wrist, right_wrist, left_hip, right_hip, left_knee, right_knee, left_ankle, right_ankle.

Positioning vs. Wikipedia and COCO

• vs. Wikipedia / Wikidata — adds machine-readable structured attributes (arrangement, penetration type, composition codes, camera framing, intensity tier) and a geometry layer that prose lists lack.
• vs. COCO / MPII — adds names and semantics to the skeleton: these keypoints are labelled with what the position is, not just where joints are.

Intended uses

• Content tagging and search facets for adult media.
• Position/taxonomy research and cross-vocabulary synonym mapping.
• Pose conditioning for text-to-image and controllable generation.

Out-of-scope / prohibited uses

• Anything involving minors — every entry documents adults only.
• Any non-consensual application.
• The dataset carries no personal data and must not be used to identify individuals.

Limitations & biases

• Keypoints are schematic templates, not observations; multi-participant registration is approximate.
• composition_code (MMF/FFM/…) describes apparent on-camera presentation only and implies nothing about identity or orientation.
• Prevalence/difficulty ratings are qualitative editorial estimates.

Citation

@misc{ethnicerotic_pose_catalog_2026,
  title        = {Erotic Pose Catalog: named positions with COCO-17 keypoint templates},
  author       = {{Ethnic Erotic}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/EthnicErotic/pose-catalog}},
  note         = {Code: \url{https://github.com/Agaveis/pose-catalog-pipeline}; CC BY 4.0}
}

@misc{pose_catalog_pipeline_2026,
  title        = {pose-catalog-pipeline: named erotic positions with structured attributes and COCO-17 keypoint templates},
  author       = {{Ethnic Erotic}},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.21439174},
  url          = {https://doi.org/10.5281/zenodo.21439174},
  note         = {Code: \url{https://github.com/Agaveis/pose-catalog-pipeline}; Apache 2.0}
}

Contact

Maintained by Ethnic Erotic — https://ethnicerotic.com/poses. Methodology: https://ethnicerotic.com/poses/methodology. Code & pipeline: https://github.com/Agaveis/pose-catalog-pipeline.