Image research and guides/SynthID·4 min read

Does JPEG compression remove SynthID?

Learn why a smaller JPEG is not proof of watermark removal, how to compare exports, and which quality changes matter before accepting a result.

By Deprint

Reviewed by Dimitry Apollonsky

Published and reviewed

JPEG compression can change an image, but it does not reliably establish that SynthID is gone. A lower file size or a lower visual detector score answers a different question. Check the exact compressed output with a compatible watermark verifier.

Compression can also damage the details you wanted to keep. A useful comparison measures both the intended file-size saving and the resulting image quality, then reports any detector result separately.

Why compression and watermark removal are different goals

A compression encoder tries to represent an image with fewer bytes. A watermark verifier tries to recognize a particular signal. Those objectives do not align automatically.

Google’s original SynthID description specifically discusses resilience to common image edits and lossy compression. That does not mean every watermark survives every possible transformation. It does mean “save as JPEG” is not a justified universal removal claim.

Same scene, different kinds of loss

Conceptual illustration. File size and detection are different measurements.

Lossless encoding

A fixed pixel array
Encode more efficiently
Same decoded values
What this means

File bytes and byte count may change without changing the picture samples.

Lossy encoding

A fixed pixel array
Discard some information
Approximate decoded values
What this means

The output can be smaller, but exact original values may no longer be recoverable.

Watermark verification

Either exported file
Run a compatible check
An observed result
What this means

Neither “small file” nor “looks identical” substitutes for a watermark check.

An image can remain visually recognizable after information is discarded. A detection signal can also remain detectable after some changes. Conversely, a visibly damaged output is not proof that the specific watermark check will be negative.

Do not confuse three different percentages

A JPEG “quality” setting, a file-size reduction and an AI detector score are separate quantities. Quality settings are encoder-specific controls, not percentages of original truth or detail retained. A detector’s displayed percentage is not the fraction of pixels that are generated.

Number What it describes What it does not prove
Encoder quality setting A setting in one export implementation Equivalent quality in every app
File-size reduction Bytes saved relative to a named baseline Watermark removal
Similarity metric Agreement under a particular comparison procedure Every important detail is intact
Detector score Output from a named classifier Universal detectability

Compare actual outputs rather than assuming that quality 80 in two applications creates the same image. Dimensions, chroma handling and metadata options can also differ.

What to inspect in the compressed image

Look at small text, sharp edges, fine hair, fabric and smooth gradients. Compression can make lettering harder to read, produce halos near edges or create visible block patterns. A whole-image view may conceal these problems.

For a simple illustrative case, consider a menu photograph with a small price in the corner. The overall scene can look nearly unchanged while a digit becomes unclear. A global similarity score can remain reassuring even though the detail that matters to the reader is damaged.

Choose those important regions before testing. Otherwise it is easy to accept a favorable detector result and only later notice that the image no longer serves its purpose.

A fair compression comparison

Start with an untouched source. Export a small set of JPEG candidates at identical dimensions while changing one encoder setting. Record the actual byte count and keep each output. If you also resize, create separate size controls so the two effects are not bundled together.

Use the same viewing procedure for all candidates. Compare at delivery size and inspect chosen detail regions at matching scale. If testing a detector, submit the exact output files and record the service and date. A screenshot of a JPEG preview is a new rendition and changes the comparison again.

For a published study, decide the quality acceptance rule before choosing the lowest score. Report outputs that failed the quality rule, not just the best-looking successful example.

What Deprint’s portrait sequence can tell us

In the selected portrait experiment, a camera-processed native intermediate had a Hive score of 7.1%. The final resized JPEG had a score of 2.9%. That difference does not isolate JPEG compression: resizing and encoding were both part of the final delivery step.

It would be inaccurate to write “JPEG removed another 4.2 percentage points of AI.” The observed score difference belongs to those exact files and that detector. It is not a quantity of AI content removed, and it is not a standalone compression experiment.

The image-quality article shows the actual before/after files. The methods describe the dimensions and comparison boundaries.

Choose compression for a measurable reason

If you need a smaller attachment or faster-loading page, compare file size against acceptable visual quality. That is a useful outcome even when watermark detection does not change.

If your real requirement is unchanged decoded pixels, choose an appropriate lossless workflow and verify it. If the requirement is a detector result, measure that result separately. No compression setting can stand in for all three decisions.

This guide reports no new JPEG removal sweep. It explains how to assess such a claim without confusing file size, image quality and detection.

Sources

  1. Google DeepMind: identifying images with SynthID
  2. Google: WebP compression techniques
  3. Deprint: methods, controls and dated observations