fileformer
search⌘KtranslateENexpand_moreupload_fileConvert
Converters8 file families · 44 direct conversions
What changes in a conversionarrow_forward
home/image-tools/image-watermark

Add Watermark to Images Free

Mark your photos before they travel without you.

Add your name, handle or copyright line to one photo or a whole batch. Choose a discreet corner mark or a tiled pattern that's hard to crop out. Images are processed in your browser and saved at full resolution.

meadow.jpg · 1800 × 1200
meadow.jpg1800×1200
text
layout
south_eastCornertitleCentregrid_onTiled
position
size4% of short side
opacity60%
colour
WhiteBlackPinkAuto
downloadSave imageSaved as JPEG at quality 92, full original resolution, with your text in EXIF and XMP. Auto picks black or white from the brightness under the mark, per image.
~/can-it-be-cropped

A corner mark survives sharing. It doesn't survive a crop.

The dashed box is the largest crop that keeps most of the photo but removes your mark completely. For the current settings, it shows how much of the image someone could keep without it.

image kept without the mark92%protectionweak
Someone could crop this edge off and keep 92% of the picture. Move the mark closer to the subject, or switch to Centre or Tiled, if that matters.
~/what-a-watermark-does

What a watermark protects, and what it doesn't.

A visible mark is about credit and deterrence, not locks. Knowing its limits helps you choose how strong to make it.

badgeworksCredit when sharedYour name travels with the picture through reposts and screenshots.
cropeasy to beatCroppingCorner marks are removed with one crop. Centre and tiled marks force someone to damage the image.
auto_fixcan be beatenAI removalModern tools can paint over a faint, simple mark. Higher opacity and tiling make that much harder.
descriptiondoesn’t decide itLegal ownershipYou own your photo’s copyright either way. A mark just makes the owner obvious.
~/same-mark-every-time

The same mark on every photo teaches software how to remove it.

A semi-transparent mark is a simple blend: published = (1 − α) × photo + α × ink. One image can’t separate the two. But average many photos carrying the identical mark and the photos blur into a smooth grey while the mark stays sharp, giving α for every pixel, which can then be divided back out. Below, your current mark is stamped on 32 generated photos and removed that way.

Photos the attacker has
Photo 1, watermarked…
Mark estimated from 32 photosα per pixel · white = opaque
Photo 1 after removal…
PhotosRemoved · same positionRemoved · moved each time
1……
2……
4……
8……
16……
32……
PSNR is measured only over the pixels the mark touches, against the unmarked photo; higher means closer to the original. Around 30 dB and up the mark is hard to see. The attacker is given the mark’s outline from photo 1 (it is visible) and estimates the background under it from the surrounding pixels of the averaged photos.
hourglass_topRunning the removal…
The attack here is deliberately simple: it averages the photos, fills in the background under the mark from the pixels around it, and assumes the ink colour is visible. Dekel et al. (CVPR 2017) use image gradients and optimisation, and remove marks almost perfectly from a few dozen stock photos. Their suggested defence is the one in the toggle: vary position, size or opacity per image. Ink set to Auto is treated as white here.
~/invisible-marks

Hiding 32 bits in your photo, then trying to break them.

An invisible watermark hides an ID inside the pixels themselves. Two classic methods are compared on your current photo, with a 32-bit ID taken from your watermark text. Least-significant-bit (LSB) hiding rewrites the lowest bit of each pixel. Spread-spectrum adds a faint ± pattern across 8 × 8 blocks, so each bit is spread over many blocks. Each marked copy then goes through eight real edits, and the ID is read back without the original.

Spread-spectrum strength±3 levelsBrightness change per block. Stronger survives more, but starts to show in flat skies.
ID (FNV-1a of your text)…
Working size…
Blocks per bit…
Spread-spectrum PSNR…
LSB PSNR…
Marked copyspread-spectrum
Difference × 16grey = unchanged
Edit applied to the marked copyLSB · wrong bitsSpread-spectrum · wrong bits
Out of 32 bits; random guessing gets about 16 wrong. JPEG edits are real browser encodes and decodes. Resized copies are scaled back to the original size before reading, as a detector would.
hourglass_topEmbedding and attacking…
~/reference

Nine ways to put your name on an image.

Visible marks live in the pixels; metadata lives beside them and disappears whenever a site or app strips it. Survival claims for the LSB and spread-spectrum rows are measured above; the rest are from general knowledge of the standards and vendor documentation.

MethodWhere it livesSurvivesLost to
Visible overlaythis toolthe pixelsRe-saves, resizing, screenshots, reposts.Cropping (corner marks), inpainting, and estimation across a batch with identical placement (lab above).
EXIF Artist / Copyrightthis tool, optionalAPP1 Exif · TIFF tags 0x013B, 0x8298Lossless copies and most desktop editors.Metadata stripping by many upload and messaging services; screenshots; ASCII only before Exif 3.0.
XMP dc:rightsthis tool, optionalAPP1 XMP packet (ISO 16684-1)Same as EXIF, with full Unicode text.Same as EXIF.
IPTC Photo MetadataXMP (IPTC Core) or legacy APP13 IIMProfessional workflows; stock and news agencies.Same as EXIF.
LSB steganographylowest bit of each pixelLossless PNG copies only.Any JPEG save, resize or brightness change (lab above).
Spread-spectrumfaint pattern spread across the pixelsJPEG, resizing, brightness (lab above).Crops and rotation without a sync pattern; anyone who knows the key.
Commercial digital watermarkspixels, with sync patternsDesigned for crops, print-and-scan and recompression.Strong distortion; reading needs the vendor’s detector.
C2PA Content Credentialssigned manifest in JUMBF (JPEG APP11)Lossless copies; tampering is detectable.Stripping, unless paired with a soft-binding watermark.
AI-image watermarkspixels, embedded by the generatorCommon edits, per vendor claims.Detector access is limited; robustness is vendor-reported.
Timeline
c. 1282Paper watermarks appear in Fabriano, Italy.
1992TIFF 6.0 includes the Copyright tag (0x8298), later inherited by Exif.
1995Digimarc is founded to commercialise digital watermarks.
1997Cox, Kilian, Leighton and Shamoon publish spread-spectrum watermarking for multimedia.
1998Exif 2.1 gives JPEG photos Artist and Copyright fields.
2001Adobe introduces XMP.
2005IPTC Core, an XMP schema for photo metadata, is released.
2012XMP becomes ISO 16684-1.
2017Dekel et al. show identical visible watermarks across a collection can be estimated and removed automatically.
2021The C2PA is formed to standardise content provenance.
2022C2PA specification 1.0 is published.
2023Exif 3.0 adds UTF-8 text; Google DeepMind announces SynthID for AI images.
Pre-digital dates are approximate; check primary sources before citing.