Guides
What an AI image carries, and what cleaning changes
An AI image can be marked in three places: the file’s metadata, the pixels, and the platform that shows it. Each guide below covers one of them, with the sources we used and the tests we ran ourselves. From image formats on, the groups are our own tests of the free tools they describe, on public files.
Content Credentials and metadata
- How to remove C2PA Content Credentials from an image
The block to delete in each format, 20 real files cleaned with sizes before and after, and two bugs we fixed.
- What is C2PA, and what do Content Credentials record?
Two real AI-image manifests read field by field, including one that records a ChatGPT image edited in a Google tool.
- How to check an image for Content Credentials
Which C2PA viewer to use, the ExifTool fields that matter, and a limit we found and fixed in our own panel.
- Is this image AI? What the file can and can’t tell you
Half of 50 AI images said nothing; C2PA survived none of nine everyday re-saves.
- C2PA vs SynthID vs platform labels: what each one is
One table: four everyday edits against the three layers that mark an AI image.
- ExifTool, a screenshot, an editor or WipeTheAI: which to use
A byte-level test of four ways to strip AI metadata, and why our own tool is not the pick for metadata alone.
Invisible watermarks
- What is SynthID, and how does the watermark work?
The pixel mark, the text mark, and the parts Google keeps private.
- Is there a SynthID detector? How to check an image
Gemini’s answer, Google’s portal and OpenAI Verify, compared, and why our label says Likely instead of detected.
- Can you remove a SynthID watermark? What we measured
Six approaches we tried against Gemini’s check, the one we kept, and what it costs the picture.
- The watermarks on Gemini and Nano Banana images
The sparkle, the pixel mark and the file record on Google AI images, one row each.
What each AI tool puts in its images
- What each AI image generator puts in its files
Seven generators in one sourced table, plus the three we could not source.
- What ChatGPT puts in the images it makes
The 2024 manifest, the 2026 watermark, and what OpenAI’s verify page looks for.
- Removing prompts and workflows from Stable Diffusion and ComfyUI images
The exact chunks and EXIF fields, the off switches, and what our remover took out of 15 real files.
- Content Credentials from Adobe Firefly and Photoshop: when they appear
The three ways Adobe attaches credentials, the forum reports, and the clause in Adobe’s terms.
Privacy and the rules
- How to remove EXIF data and GPS location from a photo
GPS, serial numbers and time zones: the tags to worry about and the free switches that remove them.
- Why a real or lightly edited photo gets an AI info label on Instagram
The metadata behind the label on edited real photos, and what Meta has and hasn’t published about it.
- Is it legal to remove AI metadata from an image?
Article 50, section 1202 and Meta’s rules read side by side, with the questions no court has answered yet.
- AI image provenance glossary
From APP11 to XMP: nineteen terms, each with a byte location or a number, and a link to the full guide.
AI video
- What Sora and Veo videos carry, and what survives a re-upload
Seven sample videos opened box by box, fifteen FFmpeg copies checked, and a lossless remover proved frame by frame.
Image formats and file size
- AVIF vs WebP vs JPG, measured at the same quality
Our equal-quality byte counts and saving times for WebP, AVIF, JPG and PNG, set against Google’s claims.
- What quality should you save a JPG at?
Measured JPG size against quality at 60, 75, 85, 95 and 100, plus how WebP and AVIF numbers differ.
- Why your PNG won’t shrink, and what to save it as instead
Four PNGs saved five ways, and why screenshots should leave PNG behind.
- Transparent backgrounds in PNG, JPG, WebP and AVIF
One transparent logo converted five ways, and why the JPG saved nothing.
- How to compress images for a website
Pixel counts versus bytes on a real photo, plus the screenshots that got bigger.
- Compress an image to 100 KB, and what you give up
How many pixels survive 50 KB, 100 KB and 1 MB limits, from six real images.
- Resizing to an exact size: crop, bars or stretch, measured
What crop to fill cuts, how big the bars get, how far stretch distorts, and why a screenshot grew.
- Animated GIF to WebP: how much smaller, and what breaks
Measured savings from GIF to WebP, when GIF wins, and why WebP back to GIF grows.
- Does converting an image remove its metadata?
Which converters kept a test photo’s GPS, and what our own tools remove.
- HEIC to JPG: what you lose, and the GPS you might not want to keep
Location, size and colour when iPhone photos become JPG or PNG, from five test files.
- SVG to PNG: why the text changes, and which size to export
Missing fonts, web fonts that never load, and file sizes at 1×, 2× and 4× for nine SVGs.
- Which favicon files a site needs in 2026
What Chromium really fetches for eleven sets of icon tags, and how big each favicon file is.
- HEIF, HEIC and HIF: what the file endings actually mean
What .heif, .heic and .hif files really contain, and which ones open where, from 69 public files.
Photo editing and privacy
- Social media image sizes for 2026, from each site’s own help
Instagram, YouTube, LinkedIn, X and Facebook sizes from their own help pages, and the limits that bite.
- Why your photo is sideways on one site and not another
Which ways of showing a photo read its rotation tag, and why WebP files stay sideways.
- How easily is a visible watermark removed?
Which watermark placement survives crops and painting-out, from six layouts on three photos.
- Can blurred faces and number plates be recovered?
Blurred and pixelated plates matched back 12 times out of 12; what that means for faces.
- Removing a background on your own device: how good is it on hair, glass and products?
Matte error on people, animals, glass and products for four free models that run without uploading.
- Batch-editing photos without uploading them: what we measured
Times, file sizes and failure handling for photo batches run on the device, laptop and iPhone-like.
PDF files
- What your PDF says about you, from ten real files
Ten PDFs opened: author codes, app versions, dates and an add-in ID nobody meant to publish.
- How much a PDF really shrinks, and when the level matters
Photos and scans shrink by up to 92%; text reports barely move, whatever level you pick.
- What survives when you merge, split or delete PDF pages
Which page tools keep bookmarks, links, forms and tags, and the hidden-page bug we fixed.
- PDF pages as JPG or PNG, and which dpi to choose
Text pages favour PNG, photo pages favour JPG, and each dpi step roughly triples the file.
- Why PDF text comes out empty, jumbled or broken
Scans, formulas, tables and hyphenation: where text extraction fails, with real examples.
- Do images come out of a PDF at full quality?
Most PDF makers store your JPEG untouched, camera details and location included; one saved it again at a third of the size.
- Photos to one PDF: page size, margins and file size
Page size barely moves the file; turned photos and pictures hidden after the photo do.
- A drawn signature vs a digital signature in a PDF
A drawn signature is a few curves on the page; a certificate signature is a checksum of the whole file.
- What turning a PDF grey changes, measured on eight files
Photo PDFs shrink most; text files barely move; and every page matched a grey rendering of the original.
- PDF passwords: what an owner password really stops
Restricted PDFs opened in ten readers: which honoured no-copy and no-print, and which just read the text.
- Black boxes don’t redact: the text we recovered from “redacted” PDFs
The Manafort filing’s black bars, our own box-only test, and what true redaction does to file size.
- OCR accuracy on photos of receipts, pages and screenshots
Character error rates for screenshots, scanned pages and real receipts, and the fixes that worked.
Markdown and web text
- Turning a web page into Markdown: what survives the trip
Links, pictures, tables and citations from one real web page, and the table fix it led to.
- Markdown to PDF: what prints well
Code lines, wide tables, long tables and GitHub-only extras on paper, with the fixes the test led to.
- Spreadsheet cells to a Markdown table, and back to text
Pipes, line breaks, tabs and semicolon files between spreadsheets, Markdown tables and plain text.
- Curly quotes, em dashes and ellipses: when plain text needs them gone
Garbled ’ characters, broken JSON and CSV, SMS limits, and counts from six books.
- Do ChatGPT answers contain hidden characters?
Counts from six OpenAI models, the o-series narrow spaces, copy and paste in Chromium, and hidden tag messages.