EthnicErotic is a catalog of the world's peoples with a phenotype layer measured by a vision model against controlled vocabularies, an erotic pose catalog built from keypoint geometry, a comparison layer for text-to-image generators, and a reader feedback loop that decides which portraits the site shows. This log explains how each part works.
This is a running log of what we build on EthnicErotic and how it works. The site is a reference catalog of the world's peoples, and it runs on pipelines: candidates harvested from open knowledge bases, profiles written from encyclopedic sources in many languages, features measured by a vision model against controlled vocabularies, positions stored as keypoint geometry, and portraits judged by readers.
Start with the systems, which lay out the machinery: what each one does, the stages it runs in order, and a figure that measures it. The dated log follows, newest first, recording each build with the technical detail behind it.
The systems
Seven systems carry the site, from the pipeline that authors new peoples to the loop that lets readers replace a portrait. Each card lists its stages in order and one figure that measures it.
New peoples enter the catalog as candidates harvested from Wikidata and leave as full profile pages. Each candidate is screened against a written rubric that separates peoples from governments, census categories and list pages. A language model then writes the profile in US English from the best available Wikipedia language edition, from French to Russian to Indonesian. Before a page publishes, it is matched to its own Wikipedia article, one article to one page. Pages publish in waves ordered by documentation depth, so the best-documented peoples go first.
Twenty-two controlled vocabularies define 196 dimensions of the face and body, each anchored to a published scale such as the Fitzpatrick skin types, and the analysis prompt, data schema, aggregation and documentation are all generated from those vocabulary files. A vision model reads real reference photographs against the dimensions a photograph can show, and a deterministic aggregator with no language model in the loop turns the readings into per-people distributions with sample sizes. A test-retest reliability study, scored with Cohen's kappa, is written back into the head and neck vocabulary files it measured.
Portraits live in per-people galleries rather than as a single picture. Each portrait prompt is built on one of four style registers, traditional, modern diaspora, modern urban or mixed, so heritage peoples appear in traditional dress and modern diaspora identities in contemporary clothes. A language model drafts the prompt from that people's homeland, phenotype profile and register, and each render from the pipeline records how it was made. Every render joins the gallery as a candidate, and readers' votes choose which portrait leads the page.
Each of the 219 positions is structured data: a 17-point keypoint template for every participant, composed deterministically from shared base-posture skeletons and a placement transform. The site draws every figure straight from those keypoints as a shaded articulated mannequin, with no image generation, and layers separate vocabularies of actions, camera moves and context on top. Positions link to their sources in historical manuals, from the Kama Sutra to the Perfumed Garden, and to museum-attested art, and a crosswalk maps them to outside vocabularies.
The people is held constant and the generator varies: five text-to-image generators draw the same people from the same prompt, and the same vision model that built the real-photo baselines reads every render. Wherever a people has a real-photo baseline, renders are scored against it, and readers vote blind on pairs, with generator names hidden until the vote. Nothing ranks peoples, and when a generator misses a people, the miss is recorded as that generator's result.
Under the catalog's portraits sits a one-tap question: does this look right? No account is needed to answer it. A "Looks off" tap opens reason chips for skin tone, face, hair or setting, an inline deck deals up to five portraits without leaving the page, and verdicts are published as open data with the surface each was cast on. When a majority of readers say a portrait looks off, new candidates are rendered, the flagged portrait is replaced, and the page says so.
The catalog's foundations are published as open research. Two datasets on Hugging Face carry the peoples, the vocabularies, notable people, image observations, reader verdicts, generator comparisons and the pose catalog under CC BY 4.0, and both pipelines are public under Apache 2.0 with methodology papers. Each pipeline has a concept DOI on Zenodo, so the work can be cited and rebuilt.
The log
When a majority of readers say a portrait looks off, the site renders replacements, promotes one and says why on the page. Twenty faces changed in a day.
On the site:Find a people and judge its portraitsHow the generator comparison worksText-to-image model catalog
Waves 4a and 4b take the catalog to 4,076 peoples pages, and no people is published under a name they reject as an insult.
A 2023 MRI study of gluteal anatomy, read against the site's own regional profiles, with fifteen group pages linked from the article.
Body-feature pages open for 120 peoples, each built on the anatomical profile of the people's homeland region and linked from its page.
On the site:The phenotype atlasHow the catalog is builtTurks: features overviewButt shape by ethnicity
Wave 2 adds 492 peoples pages, each matched to its own Wikipedia article first, and wave 3 shows 226 look-alike names are distinct peoples.
On the site:Pontic GreeksPeranakansMiamiRashaida
Five text-to-image generators, ten peoples, 200 renders scored against real-photo baselines, and a blind vote with the model names hidden.
On the site:BalochiThe text-to-image model catalogHow the comparison is scoredFive generators draw the Chechens
Reason chips turn a 'Looks off' tap into a labeled reason, and an inline deck lets readers rate up to five portraits without leaving the page.
On the site:IrishFind a people and judge its portraits
A Wikidata harvest, a written rubric and a four-step pipeline published 253 new peoples pages in one day, from the Mosuo to the Hidatsa to the Tzotzil.
115 new positions, including 30 group arrangements, each built from composed keypoint skeletons that the site draws as its diagrams.
On the site:Browse the pose catalogHow the pose catalog is builtPose catalog taxonomyThree-Man Double Team
A test-retest study over 196 portraits scored the head and neck dimensions, wrote the results into the vocabulary files, and made them citable.
On the site:About the phenotype catalogNose
2,681 portraits read by a vision model against 22 controlled vocabularies now yield 28,756 structured observations that feed the trait pages.
"Egyptian eyes" draws about 9,900 US searches a month. Trait pages answer questions like it with distributions measured from real portraits.
One-tap verdicts now land in a public table: 173 judgments across 108 groups in v5, with the labels open under CC BY 4.0.
On the site:Irish
For 61 of 104 positions, no other machine-readable vocabulary holds a precise identifier. Pose catalog v2.0 is that record, now with motion and context.
A Wikipedia backfill adds 9,792 notable people and lifts covered groups from 292 to 453. Three days later, 15,228 rows carry a photo link.
On the site:BalochiWorld atlas
The Verdict Strip puts one question under each portrait, needs no account, and hands every reader the next face nobody has judged.
On the site:IrishPose catalogLanguage-family clustersMexico
Eleven language-family hubs plus Bantu, Nilotic and Polynesian branch pages, and the open phenotype dataset rebuilt from 484 to 1,779 groups.
All 104 poses render as articulated figures computed from their keypoint data, and the pose pipeline is open source with a methodology paper and a DOI.
On the site:See the figuresPose catalog methodology
A timeline of erotic art from the Venus figurines to the lens era, and an index of 87 museum-attested works across 13 traditions.
On the site:The timeline of erotic artThe Pose CatalogStanding rear entry (bent over)Cowgirl (woman on top)
A named catalog of sexual positions with keypoint skeletons and historical provenance, published as the open CC BY 4.0 dataset EthnicErotic/pose-catalog.
A new homepage, a sitewide accessibility pass, named color variables, a mobile-first layout, 25 engraved atlas plates and one shared page header.
The Himba, the Mursi, the Omo Valley peoples, the San and the Wodaabe arrive, taking the catalog from 1,489 to 1,549 pages.
Follow any people's message board, get an alert when a new thread starts, and read your own feed of every board you follow.
On the site:WodaabeThe community boardsFind a people's board
Every portrait now starts from a prompt drafted for its group, and a style register keeps modern diaspora peoples out of historical costume.
Portraits became galleries. Members vote on every image and flag the ones that miss, and the top-rated portrait becomes each group's cover.
On the site:See the galleriesWestern AfricaEthnicErotic home
About 13,438 words grounded in peer-reviewed research: skin tone scales, hair, eye color, body composition, stature and how each one is measured.
On the site:Read the articlesHow skin tone is classifiedEvery scale the catalog usesHair texture, color and density
About 200 country pages break each nation into its peoples from cited census and genetics sources, linked by homeland, diaspora and Wikidata IDs.
22 controlled phenotype vocabularies: 196 dimensions, 853 defined terms and 113 references, with every dimension anchored to a published scale.
The code that builds the phenotype dataset is public under Apache 2.0, with a methodology paper and a Zenodo DOI, so anyone can rebuild and cite it.
Group pages gain Notable People linked to Wikipedia, and a vision model turns 5,668 reference photos into structured trait observations, each traceable.
All 484 groups, each with an original profile of 300 to 450 words, published as an open CC BY 4.0 dataset one load_dataset() call away.
On the site:Browse the catalogPhenotype Atlas
On March 29 the atlas took a scientific name and its own landing page, every group page became a phenotype profile, and the site centered on AI creators.
On the site:Open the Phenotype AtlasEyesNoseLips
Ten days in March made the atlas a conversation: boards on every people's page, public profiles, threaded replies, likes, notifications and one feed.
On the site:Join the community feedWorld atlas
The atlas of the world's peoples, first coded in 2021, moves to a modern stack with a Body Atlas, member boards and an AI portrait wizard built in.
Open research