Data cardsynthetic-face-v1 · v2026.1Available · v1

Synthetic Face
Social-Media Benchmark

Name: Synthetic Face v1
Creator: Margen
License: https://www.margensoftware.com/terms

Does your detector still catch an AI-generated face after it has been through a social feed, compressed, resized, and re-encoded, rather than arriving as a clean studio render?

6,650

Real references

19,842

Synthetic faces

Conditions

Demographic cells

SDXL + InstantID

Generator

Read the benchmarkSee how other models performed

An SDXL synthetic face, light female cell

ModalityStill images, single frontal face per image

Demographic axis6 skin tones x 2 genders = 12 cells

ClassesSDXL synthetic faces plus matched real references

ConditionsClean plus platform re-encode (compression, blur, noise, resize, pipelines)

LabelingPerceptual demographic labels by multi-model consensus, not biometric ground truth

Format.jpg, per-item demographic and condition labels

01 / The gap

Clean-render accuracy is not field accuracy.

Detectors that score well on clean synthetic faces often degrade sharply once an image has been through a platform's upload and re-encode pipeline. This benchmark isolates that gap.

02 / Coverage

Every item is labeled on a two-part axis, so performance can be read per cell rather than as a single average.

Twelve cells, read as a grid.

Skin tone

female

male

very_light

very_light · femaleITA greater than 55

very_light · maleITA greater than 55

light

light · femaleITA 41 to 55

light · maleITA 41 to 55

intermediate

intermediate · femaleITA 28 to 41

intermediate · maleITA 28 to 41

tan

tan · femaleITA 10 to 28

tan · maleITA 10 to 28

brown

brown · femaleITA -30 to 10

brown · maleITA -30 to 10

dark

dark · femaleITA less than -30

dark · maleITA less than -30

Skin tone uses a 6-band ordinal scale reconciled against Individual Typology Angle, so cells stay comparable across items. Swatches are indicative of the band, not a measured value.

03 / Conditions

Each condition emulates a platform pipeline.

Applied identically to reals and fakes. Facebook, Instagram, TikTok, and X pipelines plus a compression, blur, noise, and resize sweep.

00 · baselinecleanThe generated face with no re-encoding.

01 · compressionjpeg_q70JPEG re-encoded to quality 70.

02 · resizeresize_0.5Downscaled to half resolution.

03 · noisenoise_10Additive sensor-style noise.

04 · platformig_pipelineA full social-platform upload and re-encode.

04 / Method

How the corpus is built and labeled.

01Generated in houseSDXL + InstantID, identity-conditioned

02Consensus labelingMulti-annotator consensus

03Emulated re-encodeDesigned to match social-media conditions

ProvenanceReal reference faces are licensed from commercial stock-media providers; synthetic faces are generated in house. Each delivered item is licensed for your evaluation use.

05 / Annotation

Every item names its attack with a TTP tag.

A pass or fail count does not tell you what to fix. Each item carries a Tactic, Technique, and Procedure tag: the tactic is synthetic impersonation, the technique is identity-conditioned synthesis, and the procedure is the specific generator. Group misses by any of the three to target a fix.

TTactic, the goal (impersonation)TTechnique, the method (identity-conditioned synthesis)PProcedure, the generator (SDXL + InstantID)

Group misses by any row

06 / Use and limits

What it is for, and what it is not.

Use it for

Detector evaluation and robustness testing
Locating where a detector breaks, by cell and by condition
Training and fine-tuning as the corpus grows

Honest limits

One synthesis family (SDXL + InstantID); more generators on the roadmap
Frontal face crops; not a census of the attack surface
Demographic labels are perceptual consensus review, not biometric ground truth

07 / Samples

Synthetic faces, one per cell.

SDXL synthetic attack face, clean, light · female — cleanlight · female

SDXL synthetic attack face, clean, tan · male — cleantan · male

SDXL synthetic attack face, clean, light · male — cleanlight · male

SDXL synthetic attack face, clean, brown · female — cleanbrown · female

SDXL synthetic attack face, clean, very_light · female — cleanvery_light · female

SDXL synthetic attack face, clean, intermediate · male — cleanintermediate · male

SDXL synthetic attack face, clean, very_light · male — cleanvery_light · male

SDXL synthetic attack face, clean, dark · female — cleandark · female

SDXL synthetic attack face, clean, dark · male — cleandark · male

08 / Query

Query the corpus the way you would query a database.

Every dimension is a filter. Compose them, read a live .total, then pull only what you need.

kind

fakereal

skin_tone × gender

very_lightlightintermediate · femaletanbrowndark

condition

cleanjpeg_q70ig_pipelineresize_0.5

generator

sdxl_instantid

Live count

.total

returned for your exact filter

request

GET /v1/items
  ?benchmark=synthetic-face-v1
  &kind=fake
  &skin_tone=intermediate
  &gender=female
  &condition=clean

Built forDeepfake-detection companiesTrust and safety teamsIdentity-verification and KYC providers

Pull it and score your detector.

Access is via the Margen platform. Generate an API key, then follow the API docs to pull the catalog and measure per cell and per condition.

Get an API key API docs

Synthetic FaceSocial-Media Benchmark