Your file is carrying more than the picture.

Drop it in and press Clean. The boxes are already set the way most people need them, so you can ignore everything below.

Some things happen no matter what you tick. Every file gets its hidden characters pulled out and its AI tags wiped. There is no switch for that, because there is no good reason to skip it.

Normalise odd characters text, on by default

Some letters can be typed more than one way. A wide Ａ and a normal A are the same letter to you and different characters to a computer. Swapping a few of them around is one of the quiet ways text gets marked. Ticked, everything is rewritten the normal way. Untick it if you need the file back byte for byte, like source code.

Fix disguised letters text, on by default

The Russian о and the English o look identical and are not the same character. Change a handful in a paragraph and you have a fingerprint nobody can see, one that survives copy and paste. Untick this if your text is genuinely in Russian, Ukrainian, Bulgarian, Serbian or Greek, or real words get turned into English letters and the spelling breaks.

Keep ordinary file details images, video, audio, off by default

Photos, video and music carry a hidden information panel: camera model, the date, the software used, song titles, and often the exact spot on the map. Left unticked all of it goes. Ticked, the ordinary camera and song details stay and only the AI parts are removed. The location buried in a holiday photo is usually the thing you would least want to post in public.

Clean inside documents PDF, Word, HTML, Markdown, on by default

A document hides things in two places: the properties box where the author and title live, and the words on the page. Ticked, both are cleaned and the invisible characters between your visible words are pulled out too. There is not much reason to turn this off.

Pixel watermark removal images, usually not installed

Some AI image tools hide a pattern in the pixels themselves. Deleting the file's hidden information does nothing to it, because it is part of the picture. The only known way to disturb it is to feed the image through another model and have it redraw the picture. It is missing on a normal install because the model is a separate download of roughly ten gigabytes and needs a proper graphics card. It has nothing to do with video.

Erase a logo video, off by default

This one is different from all the others. Everything else edits hidden information and leaves your file otherwise untouched. This paints over part of the picture: you pick a corner and that rectangle is smudged over using the colours around it. The whole video is rebuilt to do it, so it is slow and the quality drops a little. It only works on a logo that sits still. If the logo drifts around the screen a fixed rectangle cannot follow it.

This is the hosted version, so your file is uploaded to the server that runs it. It is written to a temporary folder, processed for that one request and deleted straight afterwards. It is never stored, logged or passed to anyone else.

• Hidden characters: zero width spaces and joiners, Unicode tag blocks, variation selectors, direction controls and disguised letters.
• Metadata: C2PA and Content Credentials, XMP and EXIF generator fields, PNG text chunks, document properties, and the metadata boxes inside MP4, MOV and WebM.
• Pixel watermarks: optional, needs a separate model and a graphics card.

Text also gets an AI writing style score, based on sentence rhythm, word variety and stock phrases. That is a hint, not proof, and nothing is rewritten because of it.

Video, what comes out and what stays

Metadata is stripped without re-encoding. Each metadata box is overwritten in place with padding of exactly the same size, so the file keeps its byte length, every internal offset stays valid, and the picture is untouched.

• Removed: C2PA and Content Credentials manifests, XMP packets, generator and encoder tags, and the WebM equivalents.
• Not removed, a moving logo. It drifts around the frame, so a fixed rectangle cannot cover it. Following a moving mark needs frame by frame tracking, which is a different problem and is not built here.
• Not removed, SynthID. It is spread across every frame and the audio, and built to survive re-encoding, cropping and compression. There is no public way to detect or strip it. Nothing here claims to, because nothing here could.

A cleaned video is not a mark free video. The record of where it came from is gone. A moving logo or a SynthID signal in the picture is still there.

AVI, WMV, FLV, MPEG-TS and a few other formats have no reader here. They are turned away rather than processed, so nothing is ever written back for them.

Three optional components can widen what the tool reads. None is required, because the built-in readers already handle the common formats, and your file is never sent to any of them. They simply do a more thorough job on unusual files when present.

• Extended reader: reads the hidden information panel across a far wider range of file types than the built-in parsers cover on their own.
• Credentials reader: opens a Content Credentials record and describes what it claims, rather than only detecting it and taking it out.
• PDF repair: rebuilds the internal structure of awkward PDFs so they clean properly.

The tool shows which of them are available on the machine it is running on.