Compression, crops and screenshots: the three killers
Drag three sliders and watch similarity collapse. Re-encodes change bytes while the image still looks the same, crops delete evidence outright, screenshots create a brand-new bitmap.
Distinguish visual similarity from watermark survival across a real delivery route.
02Try
Run a route preset, find the quality cliff, then compare compression, cropping, and screenshot simulation.
03Prove
You can use similarity for triage without mistaking it for attribution or payload recovery.
Loading the interactive demo for "Compression, crops and screenshots: the three killers"… The lesson text below is already here. The hands-on part loads next (requires JavaScript).
Read the result correctly
Three outputs, three different claims
The labels are deliberately not interchangeable.
Visual similarityTriage signal
Likely same image is not the same as same payload.
Carrier estimateRoute-dependent
The transformation can remove trusted regions.
Payload recoverySeparate test required
Run Inspect or an extraction test on the leaked file.
Practice kit
Take the lesson outside the sample
Fictional fixtures only · downloads stay on your device
Route-test kit
A fictional original, transformed leak, delivery template, and evidence worksheet.
A plain-text checklist for preserving the exact file and context.
Look for
Similarity can support triage without proving a payload survived.
A route test must include the actual export or delivery transformation.
The evidence record is useful even when the hidden ID cannot be recovered.
All files and identities are fictional training fixtures; they are not evidence of a real leak.
Three killers, three different failures. Compression degrades, crops delete, screenshots replace. Drag the sliders and watch a real perceptual-hash score move: the same triage signal investigators use.
Read the score correctly
Above 92: likely same image, mild processing. 78-92: edited copy, heavier path. Below 78: heavily transformed or different file.
None of those name a subscriber. Similarity is triage. Attribution needs a recovered ID plus your log.
Test at your real settings
Don't test at pristine q95 if your route ships at q72. Social feeds land near 70-82 after re-encode. Test there.
Screenshots are a new bitmap: often new dimensions, new compression, and less metadata. Always prefer the downloadable attachment.
The quality cliff
Similarity often degrades gradually, then falls off a cliff somewhere around q60 to q45 for photographic content. Plot the curve above and find your image's cliff.
Fragile watermarks die well before the cliff: the photo still looks fine at q72 while the payload is already gone. A file that looks fine has not necessarily survived.
Screenshots are a new photo
A screenshot re-rasterizes at screen resolution and applies its own compression. It may also remove metadata, so it shares the visual arrangement of colors without preserving the original file's evidence.
Enable screenshot simulation above: it downsamples to 1080px first, then applies your destroyers. That ordering matters. Detail lost early cannot be recovered.
JPEG versus WebP
Same quality number, different damage: WebP's sharper quantization kills some structures JPEG keeps, and vice versa. Never assume one test covers both.
Switch the format dropdown at identical settings. The score gap is the reason your method choice must name the format, not just 'compressed'.
Takeaways
JPEG compression can destroy fragile payloads while the image still looks fine.
A crop may delete the exact regions a fragile method trusted.
Screenshots create a new bitmap and commonly remove metadata; always download the file when possible.
Check yourself
Optional, local, instant. 0/2 answered.
Similarity score is 95 after a re-encode. What does that prove?
A 10% crop drops the verdict a tier. Why?
Evidence checkpoint
Prove the triage limits
Lab first
Use the destroyer to separate file similarity from watermark survival.
The route transformation is the experiment.
Similarity can say likely same image; it cannot identify a subscriber.
Only a recovered ID matched to a clean delivery log supports attribution.
For your niche
Photographers
Client re-exports at 'high quality' still re-quantize. Test at q85, not q95: Lightroom's 85 is where gentle damage starts.
Subscription creators
Your route is the harshest: feed re-encode plus screenshot risk. Preset 'screenshot of post' is your baseline, not your worst case.
Illustrators & designers
Flat vector-like art collapses differently: edges ringing at q60 while photos look fine. Plot the curve on your own style, not the sample.
Demo transcript
Questions creators ask
Keep handyLearn glossary11 recurring terms
Carrier
The image content that offers texture or edges where a watermark can be tested.
A carrier estimate is not a survival result.
Delivery ID
An identifier for one particular delivery event or copy.
Keep it separate from the stable recipient ID.
DCT
A transform that represents image blocks as spatial-frequency coefficients.
The Learn visualization is illustrative.
LSB
Least-significant-bit hiding: changing the smallest binary bit of pixel values.
It is fragile under common re-encodes.
Metadata
Attached file information such as EXIF, GPS, software, color, and creation tags.
Metadata hygiene does not protect pixel payloads.
Payload recovery
Successfully reading the hidden identifier from the exact leaked file.
This is different from visual similarity or carrier capacity.
Re-encode
Decoding and saving an image again, often with a different format or quality setting.
It can change pixels, metadata, dimensions, and color behavior.
Recipient ID
A stable opaque identifier for the person or account receiving a copy.
Never recycle it between recipients or releases.
Release ID
An identifier for a batch, edition, or release of content.
One recipient may receive multiple releases.
Similarity
A visual or perceptual comparison suggesting whether two files resemble each other.
It is triage, not attribution.
Stable map
A validated delivery record connecting a recovered ID to one unambiguous recipient.
Duplicate IDs break attribution.
Workflow Upgrade
You found the cliff. Choose a method that clears it
Similarity falls off a cliff your eyes miss. Take that quality number into the method picker and choose a method rated for your real route.
