A metadata lab for AI images, built by Bashige Studio

For creators: metadata you did not mean to share

Plenty of people generate virtual faces, alternate selves, fantasy rooms or entire invented trips, and then post them like ordinary photographs. The picture is the point. The metadata is an accident of the toolchain.

What a generator leaves behind is rarely just a camera block. Depending on the tool, the file can carry the name of the generator, the prompt, the negative prompt, the seed, the sampler, the checkpoint hash, the whole ComfyUI or InvokeAI workflow as JSON, the author by-line, and frequently a full C2PA content credential that states in plain words that the asset is synthetic. All of it sits in the file by default and all of it can be read without decoding a single pixel.

That matters for two different reasons. The first is privacy: the trail describes how you work — which model, which prompt structure, which pipeline — and a platform that reads it can profile you far more precisely than the picture alone allows. The second is fairness: detection heuristics that only look at metadata, or that treat a declared synthetic source as a reason to down-weight or throttle a post, can punish work that is entirely harmless. Standardising the block into an ordinary camera profile removes those unnecessary fingerprints while leaving the picture exactly as you made it.

RealTag does that by rebuilding the metadata from scratch rather than deleting a few fields. A file with a coherent camera profile and camera-typical noise is not just missing a suspicious string — it stops carrying information that was never yours to publish.

For engineers: samples that exercise the real code paths

Social and mobile platforms handle photographs all day: upload, transcode, thumbnail, EXIF-driven orientation, colour profile conversion, geo-tagging, deduplication. Those code paths are written against the shape of a camera file, and they are usually tested with photographs.

Generator output is a poor substitute. A flat render with no EXIF skips the orientation branch, the EXIF parser never sees a Make/Model pair, the noise is unnaturally smooth so texture-based heuristics behave differently, and the provenance manifest triggers an entirely separate branch that a real user would never reach. The result is that a whole class of bugs stays hidden until a user uploads a genuine photograph.

RealTag turns any image into a camera-shaped sample: real Make and Model strings, an exposure that is internally consistent with the scene it claims, focal length and 35 mm equivalent, ISO, flash, lens serial, a sensor noise distribution that matches the residual statistics of a real capture, and — if you want it — a C2PA manifest so the provenance branch is exercised too. Load a batch, apply one profile to all of them and export a ZIP into your fixture folder.

How it is built

The whole site is static HTML, CSS and dependency-free JavaScript. There is no backend, no build pipeline and no analytics of your files. The TIFF/EXIF codec, the JPEG, PNG and WebP container writers, the XMP and IPTC parsers, the JUMBF/C2PA handling, the PRNU simulator and the ZIP writer are all implemented in this repository, which keeps two properties that matter for a privacy tool: your file never leaves the tab, and the behaviour of the tool is fully auditable by anyone who opens the source.

Every result is verified before you get it: the Output tab re-parses the produced file and reports which metadata survived, whether a C2PA manifest is still present, how many AI indicators remain and the measured PRNU correlation.

Honest limits

• RealTag rewrites metadata and noise statistics. It cannot change what the picture actually shows.
• Content-based detectors still analyse the pixels. No metadata tool defeats a model that looks at the image itself.
• Camera-vendor MakerNotes are not synthesised, because a malformed MakerNote is worse than a missing one.
• The PRNU simulation reproduces the shape and measurability of a sensor fingerprint, not a specific physical sensor.
• Removing a C2PA manifest does not remove a copy that a server already holds.

If you publish AI-generated work, keep the provenance you are required to declare. RealTag is for privacy hygiene and for testing, not for passing something off as a photograph of a person who does not exist.