Person ReID API · v1 beta

Turn one person crop into an appearance embedding.

A narrow, versioned contract for teams building offline similarity retrieval, media organization and human-in-the-loop review.

Successful requests only

$0.0025no retention
$0.005030-day vector retention
  • No source image stored
  • 24-hour encrypted retry cache
  • No subscription required

Quickstart

Your first embedding in one request.

Create a named key below. Send exactly one crop containing one person. Detection, tracking and multi-person splitting are intentionally outside v1.

  1. 1 Create an account and accept the API Terms.
  2. 2 Add credits and create a named key.
  3. 3 Upload a static JPEG, PNG or WebP crop.
terminal
curl https://polyreid.com/api/v1/embeddings \
  -H "Authorization: Bearer $POLYREID_API_KEY" \
  -H "Idempotency-Key: $(uuidgen)" \
  -F "[email protected]" \
  -F "retention=none" \
  -F "model=polyreid-person-reid-v1"

Immutable model contract

Portable vectors, explicit compatibility.

Store the model revision and preprocessing version beside every vector. Compare embeddings only when both values match.

Model
polyreid-person-reid-v1
Output
768 × float32
Normalization
L2 unit norm
Preprocess
person-crop-256x128-v1
Formats
JPEG · PNG · WebP
Limits
10 MB · 25 MP · static
200 application/json
{
  "id": "cm...",
  "request_id": "req_...",
  "model": "polyreid-person-reid-v1",
  "model_revision": "<sha256>",
  "preprocess_version": "person-crop-256x128-v1",
  "embedding": [0.018, -0.044, 0.091, "... 768 values"],
  "dimensions": 768,
  "dtype": "float32",
  "normalized": true,
  "retention": "none",
  "expires_at": null,
  "charged_usd": "0.002500",
  "balance_usd": "9.997500"
}

Volume calculator

Know the bill before you integrate.

Price is captured only after a valid vector is produced and, for 30-day mode, encrypted storage succeeds.

Vector retention
Estimated API spend$25.00Taxes, when applicable, are added to credit deposits.

Stable failures

Every error includes a request_id.

401API_KEY_REQUIRED / INVALID_API_KEY

Create, rotate or replace the Bearer key.

402INSUFFICIENT_CREDITS

Add prepaid credit; failed requests are not charged.

409IDEMPOTENCY_KEY_REUSED

Use a new key when the payload or options change.

413IMAGE_TOO_LARGE

Stay below 10 MB and 25 megapixels.

415UNSUPPORTED_IMAGE_TYPE

Send an actual JPEG, PNG or WebP file.

422INVALID_IMAGE

Send one valid, static person crop.

429RATE_LIMIT_EXCEEDED

Back off using the response rate-limit headers.

503INFERENCE_UNAVAILABLE

Retry safely with the same Idempotency-Key.

504INFERENCE_TIMEOUT

Retry safely; the timed-out attempt is not charged.

Privacy, versioning and use

A production boundary you can explain.

No source-image persistence

Image bytes exist only in request memory and a bounded in-memory GPU micro-batch. They are not written to PostgreSQL, Redis, B2, logs or Sentry.

Idempotence without double spend

An encrypted response cache lives for 24 hours. Retrying the same payload with the same key returns the original response and charge.

Optional 30-day vector storage

Only the encrypted embedding and technical model metadata are retained. Delete it early at any time; expiration blocks access immediately.

Lawful commercial use

You must have rights, notices and a lawful basis for the images. Outputs are probabilistic review candidates, never proof of identity.

Version stability

V1 is never silently retargeted. An incompatible model becomes v2; announced deprecated versions remain available for at least 12 months.

Beta operating target

No contractual SLA during beta. The internal target is 99.5% availability, warm P95 ≤ 3 seconds and cold P95 ≤ 60 seconds.

Responsible use, biometrics and EU AI Act

Define the deployment before relying on the output.

Assessment statusEU AI Act classification under legal assessmentLast reviewed 2026-07-24
01

Intended purpose

Match individual person crops inside offline photo workflows for similarity retrieval and trained human review.

02

Not an identity decision

V1 does not detect people, recognise faces, assign names, track people live or make automated decisions.

03

Human oversight

Validate thresholds and error rates on your own domain. Give a trained reviewer enough context to reject or correct a match.

04

Prohibited boundaries

No unlawful biometric use, sensitive-attribute categorisation, prohibited emotion recognition, social scoring or prohibited law-enforcement use.

No conformity claim

Classification depends on the intended purpose, integration and actual use. This public information does not replace risk management, technical documentation, post-market monitoring or a required conformity assessment.

Developer console

Keys, balance and evidence in one place.

Private account data below is never rendered into the public SEO document.

Checking your developer account…