A record that lives outside the file.

Provenance has to survive the internet.

A digital image travels by being reposted, cropped, re-encoded, filtered, and passed across platforms that each rewrite the file. Unauthorized use falls into three broad categories: direct reposts (the same file, a new account, a new name), derivative edits (crops, flips, filters, collages, overlaid watermarks), and AI-generated derivatives (style imitation, image-to-image variants, models fine-tuned on a specific artist).

The established defenses all try to mark the file itself, and the file is exactly what keeps changing. Visible watermarks get cropped out. Credentials embedded in a file's metadata are stripped by a single re-save or screenshot. Platform-specific badges do not travel to the next platform. Reverse image search only helps if you already suspect a match and think to go looking.

Keep the record, not the mark.

Lantern does not depend on anything staying attached to the image. Rather than writing a mark into the file, it derives a set of fingerprints from the image and stores them in a public registry alongside a timestamp that no single party can quietly change. When a copy surfaces later, its fingerprints are recomputed and matched back to the original registration, even if the copy has been cropped, re-encoded, or reposted under a different name.

Registration: several fingerprints, not one.

A registration stores an exact cryptographic hash of the file (SHA-256), perceptual hashes that tolerate re-encoding and small edits, and neural embeddings from established vision models: a copy-detection embedding tuned to recognize altered copies, and self-supervised embeddings that capture the visual content itself. No single fingerprint is sufficient alone. Together they cover a wide range of transformations while keeping false matches rare.

Matching: exact first, then fuzzier, then by region.

A query image is checked against the registry in stages. An exact hash match is the strongest and cheapest signal. When there is no exact match, perceptual and neural signals are combined into a single confidence score that recovers re-encodes, crops, and filtered edits. A region stage matches sub-parts of an image, so a corner or a detail lifted from a registered work can still be traced back to it. Every result carries a plainly stated confidence level.

The ledger: independent and publicly readable.

Each registration is anchored to a public blockchain ledger. The record is timestamped, tamper-evident, and readable by anyone, independently of Lantern the company. If Lantern were to disappear tomorrow, the registrations would still exist and any reader could still confirm what they contain. Creators never touch a wallet, a key, or a token: the ledger is infrastructure, kept out of the way.

Built for the provenance rules now taking effect.

Content-provenance requirements are moving from voluntary to mandatory. The European Union's AI Act, Article 50, becomes fully enforceable on August 2, 2026: providers of generative systems serving EU users must mark AI-generated content with machine-readable credentials, offer public means of detection with confidence indicators, honor creator opt-outs from training data, and use more than one technique rather than relying on a single mark. California's labeling and platform-detection laws follow on parallel timelines.

Lantern's design maps onto that shape: durable provenance that does not rest on a single strippable mark, a public and confidence-scored way to check an image, and a registry that can publish creator opt-outs at stable, machine-readable addresses. It is meant to complement content-credential standards that travel inside the file, not replace them. When such a credential survives, it is a strong signal; when it has been stripped, the registry still has an answer.

Artists keep control.

You register, you link only the accounts you choose to link, and you can withdraw a registration if you change your mind. Lantern is the caretaker of the registry, not the owner of the work. Every action is recorded, and every action is yours.

Credit, never suspicion.

Lantern marks what it can verify and stays silent about everything else. The absence of a record means there is nothing to say yet, not that a work is fake. Treating no-signal as a signal would turn a provenance tool into an accusation tool, which it is not.

Authorship does not require identity.

A work can be registered without linking any social account. Linking adds trust by connecting a registration to an artist a viewer already follows, but it is never required. Sharing a legal name and keeping the rest of your life private are equally valid, and both are supported.