Questions, answered plainly

It is in pre-release and has not been published yet. Packaging is done and the app is heading for the Microsoft Store. The availability page will carry the store link the moment it is live — that will be the only place a legitimate download comes from.

Pricing has not been announced yet. That is a genuine “not decided”, not a dodge — rather than guess here, the availability page will state the terms when they are final.

Windows 10 version 1809 (build 10.0.17763) or later, including Windows 11, on both x64 and Arm64 devices. Everything it needs is bundled into the package, so there is nothing else to install first.

The codebase is cross-platform by design, but the released product targets Windows through the Microsoft Store, and no macOS or Linux release is announced. Treat anything beyond Windows as unconfirmed.

That has not been announced, so this page will not claim it. Which is worth saying explicitly, because “open source” is a claim people reasonably rely on when deciding whether to trust a tool with sensitive files — and it should not be made loosely.

No. The application contains no HTTP client and no telemetry code. Image analysis, AI detection, hashing, steganography and reporting all run on your machine.

The only outbound URLs the product can produce are optional OpenStreetMap and Google Maps references written into a geo-tag report. Those are contacted only if you click them — the app itself never requests them. The privacy page lists them in full.

Yes, completely. Disconnecting the network changes nothing about how the app behaves, because none of its features depend on a remote service. That also makes it easy to verify the claim yourself: pull the cable and use it.

In one folder under your own Windows profile: %LocalAppData%\StegoPixel. That holds your settings, the SQLite library index (stegopixel.db), saved metadata templates and local diagnostic logs. Nothing is written to the install directory and nothing is synchronised anywhere.

Because it is an ordinary folder, it is covered by whatever backup and profile encryption you already use — and deleting it resets the app, with no server-side copy left behind.

Not unless you explicitly apply changes. “Save a copy” is the default behaviour, backups are taken before destructive operations, deletions go to the Recycle Bin rather than being erased, and stripping all metadata from a set of files asks for confirmation first.

The practice of hiding data within another file, so that the existence of the data is not obvious. StegoPixel can both create hidden payloads and detect them — in image pixels and in metadata structures.

They fail in opposite ways, so the answer depends on what you are defending against:

• Pixels survive metadata stripping — but any lossy re-encode destroys them. Re-saving a JPEG, resizing, or uploading to a platform that recompresses will lose the payload.
• Metadata channels hold more, and survive pixel-level recompression — but only exist in formats that support XMP, and vanish the moment somebody cleans the file's metadata.

If the file will be edited or re-shared, use a lossless carrier such as PNG, TIFF or BMP with pixel embedding. If it will be transmitted intact and you need capacity, a metadata channel is the better fit.

Pixel capacity scales with resolution, the number of channels and the bit depth you select — more bits per channel means more capacity and more visible distortion. A metadata channel targets 44,984 bytes in JPEG, PNG, TIFF and WebP. That is deliberately a typical maximum rather than a guaranteed ceiling: external EXIF readers may stop honouring metadata profiles beyond 64 KB, and the real figure also depends on the image and the tags already present.

The number to trust is the one the app shows you: it reports real available capacity for the specific file and settings before you encode, so a payload is never silently dropped.

Encrypted payloads use AES-256-CBC with the key derived from your passphrase through PBKDF2 with SHA-256 at 100,000 iterations. These are standard, widely reviewed primitives — the same class of construction used by mainstream password managers.

Two honest caveats. First, passphrase strength still matters: a high iteration count raises the cost of guessing but cannot rescue a weak passphrase. Second, steganography and encryption solve different problems — steganography hides that a payload exists, encryption protects its contents if the carrier is discovered. Use both.

Yes — that is half of what the product is for. It analyses LSB patterns in the pixels, looks for anomalous metadata structures, checks C2PA provenance chains, and scans for AI-generation signatures. It can also report whether a payload is present without extracting it, and decode payloads when you have the passphrase.

It combines seven offline signals and resolves them into a tiered verdict — Natural, Inconclusive or AI-generated — shown alongside the score and the signals that contributed:

• Generator and editing-tool signatures left in EXIF and XMP — a table of roughly 30 known tools, each with its own confidence weighting — plus the IPTC digitalSourceType field.
• A plausibility audit of the EXIF block: a Make/Model brand mismatch, impossible or future dates, out-of-range numeric values, an all-zero GPS position, or a camera identity that carries no capture parameters.
• A stripped-metadata check for a JPEG, WebP or TIFF with a completely empty metadata block. It is deliberately the weakest signal — routine re-encoding and privacy scrubbing look identical — so it can only corroborate.
• Pixel statistics: DCT, noise uniformity, colour-filter-array and spectral measurements, blended into one score. These still work after metadata has been stripped.
• A frequency-domain scan for repeating watermark carriers. This is a heuristic pattern match against a small reference set, not a read of any vendor's production watermark key, and the app's own interface says as much.
• A decoder for the fixed invisible watermark the original Stable Diffusion pipeline embedded — a known 136-bit payload read back from the wavelet domain by majority vote. A strong match is checkable evidence; a miss proves nothing, since resizing and cropping destroy the watermark.
• Parsing of any C2PA Content Credentials in the file — the provenance claim is surfaced, not cryptographically validated.

The verdict also records which checks ran, were skipped or failed — a check that could not complete is reported as incomplete, never passed off as clean.

To be clear about what is under the bonnet: these are hand-written statistical heuristics, not a trained neural network. There is no model file and no labelled dataset. That is exactly why the app reports confidence tiers instead of a yes/no answer.

“Inconclusive” is a real outcome rather than a failure. An offline detector that never admits uncertainty is not being straight with you.

No, and you should be suspicious of any offline tool that claims it can. Every detector of this kind produces both false positives and false negatives. Metadata can be forged, stripped, or added innocently by an editor. A C2PA manifest proves what a manifest says, not that a photograph depicts reality.

Use the contributing signals rather than the headline label, and do not present a score as a determination about a specific person or image.

StegoPixel exposes 19 typed writable EXIF tags — eight strings, three date/time, five numeric and two enumerated. The eight strings are the ones people reach for most: ImageDescription, UserComment, Artist, Copyright, DocumentName, Software, Make and Model. Custom key/value pairs are written as XMP in a dedicated urn:stegopixel:custom namespace.

The whole of the metadata is readable and exportable regardless — the restriction is on writing, because most image formats cannot reliably be written back beyond a small subset of tags.

JPEG, PNG, TIFF, WebP, GIF and BMP are decoded end-to-end. Some capabilities are format-dependent:

• Metadata channels need XMP support, so they work on JPEG, PNG, TIFF and WebP — not GIF or BMP.
• Writing EXIF tags is likewise limited to JPEG, PNG, TIFF and WebP. GIF and BMP cannot hold EXIF, so the payload would be dropped.
• Pixel embedding applies to any format the app can decode.

HEIC/HEIF, AVIF and camera RAW formats (CR2, CR3, NEF, ARW, RAF, DNG, ORF, RW2, PEF, SRW) appear in the format catalogue but cannot be opened — the decoder handles the six raster formats above only. Unsupported files are reported as unsupported rather than partially analysed.

The output is designed to be presentable — forensic reports export to HTML or PDF with the result imagery, verdicts and file digests (SHA-256 and SHA-512) attached, and that is a reasonable basis for internal review or a professional write-up.

Whether it is admissible in a legal proceeding depends on your jurisdiction's rules on tool validation and expert testimony, which is a question for a lawyer rather than a marketing page.

No. A pipeline is built by picking nodes from a searchable list and ordering them. There are 45 node types and 54 ready-made presets, so most tasks start from something that already exists and get adjusted rather than built from nothing.

If you would rather answer questions than wire nodes together, the guided wizard walks you through the same pipeline one decision at a time. See the workflows page for the full node catalogue.

The defaults are built around recoverability: cleaning writes a copy rather than overwriting, backups are taken before destructive steps, deletions go to the Recycle Bin, and permanent deletion has to be explicitly requested. Metadata edits support undo while you are still in the session.

What cannot be undone is a pipeline that has already permanently overwritten or deleted files — which is why running a new pipeline against a copy folder first is worth the two minutes.

Twelve: English, German, Spanish, French, Italian, Japanese, Polish, Portuguese, Russian, Simplified Chinese, Hindi and Arabic. You can switch at runtime without restarting the app, and Arabic switches the layout to right-to-left.

Yes, and both themes are designed as a matched pair rather than one being the other with inverted colours. The app defaults to the dark theme; the accent colour shifts from cyan on dark to a deeper teal on light so it stays legible against a pale background.

See the contact section on the About page. Concrete detail helps: the app version, what you did, what you expected, and what happened instead. If it involves a file, say which format it was, since a great many surprising behaviours turn out to be format limitations.