Where Our Data Comes From, and What's Missing

Data in the starlikeness index comes from cover images and video stills published on 106 public sites, plus performer metadata parsed from public profiles. Representative images come from only four hosts, one of which supplies 2,246 of 2,333 records, so coverage reflects aggregator exposure.

Last updated Mon Aug 03 2026 00:00:00 GMT+0000 (Coordinated Universal Time)

Our data comes from two streams that meet in the middle. Face vectors are computed from cover images and video stills published across 106 public sites. Performer metadata — names, translations, biographical fields — is parsed from public profile pages. Everything we publish is derived from one or both, and where they disagree or fall silent, we have a gap rather than an answer.

The most important thing to know about the sources is how narrow they are. 2,246 of our 2,333 identity records take their representative image from a single site, and across the whole index only four distinct hosts supply representative images at all.

What are the two data streams, precisely?

One stream is computed by us; the other is inherited from publishers. They fail in completely different ways.

Face vectors Performer metadata
Origin Images published on 106 indexed sites Public profile pages
Produced by Our own pipeline (YuNet, ArcFace) Third-party editors
Completeness Near-total for indexed images 23%–89% by field
Failure mode Detection misses; bad angles, occlusion, small faces Fields left blank, placeholder strings, contradictory values
Can we improve it? Yes — index more sources Only if a source publishes it

The asymmetry matters. When face data is missing it is because we did not index the image. When metadata is missing, nobody wrote it down in the first place, and there is no amount of engineering that recovers it.

Which sources dominate, and how badly?

One site dominates almost completely, and we would rather publish that than let "106 sites" imply breadth we do not have.

Representative image host Identity records Share
Primary source (jable) 2,246 96.3%
Second host 67 2.9%
Third host 14 0.6%
Fourth host 6 0.3%

Source: our index, 2026-08 snapshot, n=2,333 identity records.

The 106 figure describes the sites contributing to the face index — the 241,792 vectors. The four hosts above describe where each identity's representative image comes from. Both are true; only the second one tells you how concentrated the identity layer is.

There is a second concentration on top of the first. Of the 106 indexed sites, only around 13 carry Japanese content IDs. The rest are largely international tube and aggregator platforms. Our view of Japanese adult video is therefore mediated almost entirely through international republication.

What is missing, field by field?

Between 11% and 77% of records, depending on the field. Here is the full accounting rather than the flattering subset.

Field Present Missing Coverage of n=2,333
Similar-performer list 2,070 263 89%
Cup size 1,375 958 59%
Debut date 1,371 962 59%
Height 1,340 993 57%
Measurements 1,316 1,017 56%
Birth date 1,301 1,032 56%
Birthplace 716 1,617 31%
Blood type 547 1,786 23%

Source: our index, 2026-08 snapshot, n=2,333 identity records.

Three notes we consider part of the data, not caveats to it:

  • The missing are not random. Records with full metadata are the best-documented performers, which usually means the most prominent. Any statistic computed over the present values describes prominent, internationally circulated performers.
  • 626 records hold a name and nothing else — no biographical field at all. They are counted in the denominator above, which is why every percentage is lower than it would be if we quietly dropped them.
  • Some earlier figures on this site were too high. Placeholder strings in the source data were being counted as real values, which inflated several coverage numbers and, worse, was being emitted into structured data on performer pages. Those values were removed. The table above is the corrected accounting and the previous numbers should not be cited.

What can this data never tell you?

It can never tell you anything about titles, studios, or release volume, because it does not contain them.

The index stores faces, not works. There is no release-date field, no studio or label field, and no title record. That rules out an entire family of statistics that would be easy to fake and impossible to support: annual production volume, studio market share, format-adoption timelines, genre shifts by year.

It also bounds the language we use. We never write that a studio released N titles. We write that we have indexed N — because what we counted was aggregator republication, and the gap between those two statements is exactly the gap between our data and the industry.

Data and method

Source. Cover images and video stills published on 106 public sites; public performer profile pages for metadata; Wikidata identifiers where a confident mapping exists.

Sample. 241,792 face vectors; 2,333 identity records (1,707 with a full profile, 626 name-string only); 13,420 similar-performer references, of which 7,004 point to unnamed identities.

Method. Faces detected with YuNet, aligned from five landmarks, encoded with ArcFace into 512-dimension vectors, retrieved by cosine similarity with a 0.40 same-person threshold. Metadata is parsed from public profiles, validated for type and format, and left empty when validation fails rather than substituted. Similar-performer lists are computed from vectors, not inherited.

Snapshot date. 2026-08.

What is not stored. Video files. Original user uploads — detection runs in the visitor's browser, only a cropped face region and its five landmarks are transmitted for matching, and query results are not retained. The index is built from published material, never from what people search for.

Known bias.

  • Extreme source concentration. 96.3% of representative images come from one host; four hosts account for all of them.
  • Aggregator mediation. Only around 13 of 106 indexed sites carry Japanese content IDs, so the material we see is what circulates internationally.
  • Documentation skew. Present metadata over-represents prominent performers.
  • Consequence. These figures measure exposure on the sites we index, not release volume or industry composition. Treat every number here as a statement about our sources first and about the industry only by inference.

Citation.

starlikeness (2026). "Where Our Data Comes From, and What's Missing."
Face index, 2026-08 snapshot (n=241,792 faces; 2,333 identity records;
106 sites; 4 representative-image hosts).
https://starlikeness.com/en/posts/where-our-data-comes-from

Related questions


Everything above describes the index, not you: face detection runs in your browser, only a cropped face region is sent for matching, and nothing about your search is written back into this data.

Frequently asked

What does the index actually store?
Face vectors, a representative image reference with face coordinates, and whatever public profile metadata could be parsed — name, translations, and biographical fields where available. It does not store video files or original uploads.
Does the index store images that users upload?
No. Detection runs in the user's browser; only a cropped face region and its landmarks are sent for matching, and results are not retained. The index is built from published imagery, not from queries.
Why do some performers have almost no data?
Because their public profiles are thin or absent on the sources we parse. 626 of our 2,333 identity records hold a name string and nothing else — no birth date, no measurements, no debut date.