Fingerprints

A fingerprint is a compact description of what a published file looks or sounds like. Every recipe that decodes the asset computes fingerprints of the bytes it publishes and keeps them on our servers, scoped to your account. When a copy arrives at verification without a usable manifest, recovery compares the copy’s fingerprints with yours to find the original; see Lookup and recovery. A recovered match is reported as recovery, never as cryptographic validity.

Only one fingerprint ever enters a manifest: ISCC Image-Code, which is the image soft binding the C2PA manifest carries. Every other fingerprint is a server-side recovery signal. It is not written into the manifest, not returned by any public response and not shared between accounts.

Choosing a set

POST /v1/sign takes an optional fingerprint_algorithms list; the rules are in Sign. In short:

• Omit it for the default set of the asset’s media type.
• A list replaces that set. ISCC Image-Code is always added for images, and at least one fingerprint is required, so recovery is never switched off.
• Only a set that differs from the default is priced, as one fingerprint line; see Retention and pricing.
• A job keeps the set it was admitted with. A later change of default never reaches it.

GET /v1/capabilities lists, per media type, the ids your deployment offers, which are in the default set and which are locked on, with a one-line note on each. Structural and sidecar containers are not decoded, so they list none and refuse the field.

What each fingerprint survives

The notes come from our 2026 fingerprint benchmark, which transformed a fixed corpus (re-encoding, resizing, cropping, overlays, rotation, pitch and tempo changes) and measured how often each fingerprint still found the original while keeping false matches rare.

Id	Media	Good at	Weak at
iscc-image	image	JPEG, resizing, colour changes, added borders	crops, rotation, overlays
dinov2	image	crops and overlays; the most robust image descriptor we run	nothing measured stands out
dinov2-lsh256	image	the same edits as dinov2, as a compact keyed 256-bit hash	slightly weaker than dinov2
pdq	image	flips and 90-degree rotations; widely used for hash sharing	crops
chromaprint	audio	codecs, noise, re-recording	pitch or tempo change
audfprint	audio	loud background speech or music; re-recording	the first 10 minutes only
iscc-audio	audio	whole recordings, as a compact open-standard code	excerpts
tmk-pdqf	video	re-encoding of the whole clip	a border crop of about 10%
vpdq	video	partial clips, frame by frame	a border crop of about 10%

The image default is iscc-image with dinov2. Audio and video keep chromaprint and tmk-pdqf as their defaults; audfprint, iscc-audio and vpdq are computed only when named. ISCC Audio-Code is kept as a recovery signal only; it is not written into the manifest.

dinov2-lsh256 projects the DINOv2 descriptor through a random matrix derived from a key held for your account alone, so its bits are meaningless under another account’s key. Changing that key would be a new calibration, recorded as such; bindings made under the old one are never compared with the new.

Catalog

The tables below are generated from the capability snapshot the API is built with. A deployment can enable fewer, so GET /v1/capabilities stays the authority at run time.

Defaults

Modality	Default (field omitted)	Always added	Selectable
audio	chromaprint	none	chromaprint, audfprint, iscc-audio
image	iscc-image, dinov2	iscc-image	iscc-image, dinov2, pdq, dinov2-lsh256
video	tmk-pdqf	none	tmk-pdqf, vpdq

Media types that keep fingerprints

Media type	Modality	Default	Selectable
audio/mp4	audio	chromaprint	chromaprint, audfprint, iscc-audio
audio/mpeg	audio	chromaprint	chromaprint, audfprint, iscc-audio
audio/wav	audio	chromaprint	chromaprint, audfprint, iscc-audio
image/avif	image	iscc-image, dinov2	iscc-image, dinov2, pdq, dinov2-lsh256
image/jpeg	image	iscc-image, dinov2	iscc-image, dinov2, pdq, dinov2-lsh256
image/png	image	iscc-image, dinov2	iscc-image, dinov2, pdq, dinov2-lsh256
image/tiff	image	iscc-image, dinov2	iscc-image, dinov2, pdq, dinov2-lsh256
image/webp	image	iscc-image, dinov2	iscc-image, dinov2, pdq, dinov2-lsh256
video/mp4	video	tmk-pdqf	tmk-pdqf, vpdq

Algorithms

Fingerprint	Name	Modality	In the manifest
audfprint	audfprint landmarks	audio	No (server-side recovery signal)
chromaprint	Chromaprint	audio	No (server-side recovery signal)
dinov2	DINOv2-S descriptor	image	No (server-side recovery signal)
dinov2-lsh256	DINOv2-S keyed LSH (256 bit)	image	No (server-side recovery signal)
iscc-audio	ISCC Audio-Code (256 bit)	audio	No (server-side recovery signal)
iscc-image	ISCC Image-Code	image	Yes (C2PA soft binding)
pdq	PDQ	image	No (server-side recovery signal)
tmk-pdqf	TMK+PDQF	video	No (server-side recovery signal)
vpdq	vPDQ	video	No (server-side recovery signal)

Licences

The fingerprints added to the choice are under permissive licences: DINOv2 code and weights (Apache-2.0), PDQ and vPDQ (BSD, from Meta’s ThreatExchange), audfprint (MIT) and ISCC Audio-Code (Apache-2.0). The ones every job already computed keep theirs: ISCC (Apache-2.0), TMK+PDQF (BSD) and Chromaprint (MIT core; the fpcalc binary we run is LGPL-2.1). ISCC Video-Code is not offered: the FFmpeg filter it depends on is GPL-licensed. ISCC’s bitarray dependency is under the Python Software Foundation licence (PSF-2.0).

Jobs admitted before the image default changed

Image jobs admitted before DINOv2 replaced SSCD in the image default computed SSCD. SSCD has since been stopped entirely (licence decision, 2026-09-28): nothing computes it, recovery never searches it, and its model is no longer part of the service. Those jobs stay recoverable by the exact SHA-256, their watermark and their ISCC Image-Code. Their stored SSCD bindings are kept only until they expire with their outputs.