Image research and guides/SynthID·4 min read

Does cropping or resizing an image remove SynthID?

Understand what cropping and resizing change, how to compare them fairly, and why a smaller image is not a verified watermark-free image.

By Deprint

Reviewed by Dimitry Apollonsky

Published and reviewed

Cropping and resizing can change an image, but neither operation guarantees that SynthID becomes undetectable. A crop removes part of the scene. A resize changes how the remaining scene is sampled. The only supported verdict is the one observed for the exact output in a compatible check.

Before trying either, decide whether you can afford to lose framing or fine detail. A result that no longer contains the information you need may not be useful even if a detector responds differently.

A crop and a resize are not the same edit

Cropping chooses a smaller rectangle. If no resampling or re-encoding is added, retained image samples can remain unchanged. Resizing changes the dimensions of the whole selected picture, usually estimating new samples from the old ones.

An editor often combines these operations. For example, selecting a square crop and exporting it at a preset social size both removes content and resamples what remains. The export may then add JPEG compression.

Cropping and resizing ask different things of the image

Illustrative geometry. These examples are not measured removal recipes.

Crop

A large scene
Keep a smaller rectangle
Less of the scene
What this means

Pixels outside the crop are gone; retained pixels may remain unchanged before export.

Resize

The whole scene
Resample to fewer pixels
Same framing, fewer samples
What this means

Interpolation combines or estimates values. Small text and thin lines can change.

Crop and resize

A selected rectangle
Scale it for delivery
Both effects together
What this means

A fair experiment needs separate controls if you want to know which step mattered.

The distinction matters because a detector can react to any of those changes. A single before-and-after pair cannot tell you which operation caused the difference.

A simple geometry example

Consider a hypothetical 2,000 by 1,000 image. Cropping it to a centered 1,000 by 1,000 square removes half its area. Resizing the original to 1,000 by 500 retains the framing but reduces the number of samples. Cropping the square and then resizing it to 500 by 500 combines both effects.

None of those dimensions is a removal recommendation. They simply show how different the inputs become. A small sign near the edge might disappear in the crop. A sign retained in the resized image might become unreadable. Either result can undermine the purpose of the picture.

Operation Framing Sample count A quality risk to inspect
Crop only Changes Usually decreases Missing context or cut-off objects
Resize only Usually retained Changes Small text and fine lines
Crop then resize Changes Changes again Both lost context and reduced detail
JPEG export afterward Usually retained Dimensions can stay fixed Encoding artifacts

Why keeping the center is not a general solution

An invisible watermark should not be treated as a visible sticker with an obvious location. Removing a corner logo establishes only that the corner logo is no longer visible. It does not demonstrate where another signal was carried or whether enough of that signal remains.

Google describes SynthID as designed to tolerate common modifications, including cropping. That provides a reason to test rather than assume. It is not evidence that every crop must produce the same result.

How to isolate the effects

Use an unchanged original and create separate branches: a crop-only candidate, a resize-only candidate and a combined candidate. Keep export settings consistent where possible. If the final application requires a specific format, add a matched export control.

For a published result, record the rectangle, dimensions and resampling implementation. “Resized in an editor” is too vague to reproduce precisely. Also record the original’s verification result and any check that failed to complete.

Choose the crop before looking at detector responses if your aim is to evaluate a normal editing workflow. Searching many crops and showing only the best one answers a different question: whether a favorable case could be found among those attempts.

What the existing Deprint controls show

In the frozen desk experiment, a size-matched control at 1,360 by 1,015 remained at 99% in Sightengine, matching the original’s displayed score. In the bakery experiment, a size-matched control at 1,435 by 963 remained at 99.9% in Hive.

Those are visual classifier observations on two scenes. They are not a SynthID resizing success rate, and they do not prove that resizing never affects a detector. They show that simple dimension matching did not explain a general bypass in those checks. The methods keep the detector and file distinctions explicit.

A sensible editing decision

Crop to improve composition or remove unwanted visible content. Resize to fit a real delivery requirement. Keep a full-resolution original so the loss is reversible at the workflow level.

If detection matters, check the final delivered rendition after all edits. If visual quality matters, inspect the details required for its use. Do not treat a smaller rectangle, lower resolution or vanished logo as a substitute for either check.

Sources

  1. Google DeepMind: SynthID
  2. Deprint: methods, controls and dated observations