Converting an image to PNG, JPG or WebP is not, by itself, evidence of SynthID removal. The outcome depends on what the conversion did to the image and on the verifier’s result for the exported file. A new extension does not identify a new provenance.
Some conversions preserve decoded samples. Others change color, transparency, dimensions or compression. First establish which operation occurred; then test the property you care about.
Rename, repackage and re-encode are different operations
Renaming picture.png to picture.jpg does not make it a JPEG. It changes the filename. A real conversion reads the source and writes a representation supported by the destination format.
A tool might preserve the image samples while changing the container. It might also decode and re-encode with lossy settings. Some tools resize automatically. These are materially different workflows even if they all have a button labelled “Convert.”
Illustration of three operations often confused with one another.
Rename
What this means
The extension changed. The image has not been JPEG-encoded.
Lossless conversion
What this means
Check bit depth, alpha and color handling. A container change does not inherently erase a pixel signal.
Lossy conversion
What this means
This creates a new rendition. Its detector result must be measured rather than inferred.
A watermark carried by image content does not disappear merely because its container has a new name. If a conversion changes detectability, report the exact conversion and observation rather than treating the destination format as a universal answer.
What the format tells you
| Format or mode | Main consideration | A useful verification question |
|---|---|---|
| PNG | Lossless storage of supported image samples | Did the tool preserve bit depth, colors and transparency? |
| Common JPEG export | Usually a lossy photo-delivery workflow | Which dimensions and encoder settings were used? |
| Lossless WebP | Lossless coding with its supported representation | Were the relevant decoded values preserved? |
| Lossy WebP | Approximate image coding, with transparency support | What changed at the chosen quality setting? |
A format’s capability does not guarantee the application used it without additional transformations. Exporting a 16-bit image into an 8-bit workflow, for example, can change information before a lossless encoder stores the result.
A transparent-logo example
Imagine a transparent logo placed on a web page. Saving it to a common JPEG workflow requires a background because that workflow does not preserve alpha transparency. A converter might choose white. The result can look fine on a white page and show a visible box on a dark page.
This example has nothing to do with proving watermark removal. It shows why “the image looks the same in my preview” is not a sufficient conversion check. The preview background can conceal a real change.
For a photograph, the hidden difference might instead be a color conversion or a reduction in bit depth. Inspect the image in its intended context, not just inside the converter.
Can a lossless conversion remove a pixel watermark?
If the relevant decoded image samples are exactly preserved, you should not credit the conversion with intentionally removing a signal represented in those samples. A verification service could still behave differently because it handles formats differently. That would be a property of the full submission and verification path, not proof that changing an extension erased the watermark.
This is why a careful report distinguishes image preservation from service behavior. Record both the file format and the actual verifier response. If two formats produce different answers, repeat the check with controlled encodes and preserve the exact files.
Can a lossy conversion change the result?
It can change the input that a detector receives, so a changed result is possible. But possibility is not reliability. A method that works on one picture can fail on another, and a change that lowers a visual classifier score can leave a watermark detectable.
Test the original first. Then check the converted file in the same relevant interface. If your conversion also resized the image or removed metadata, isolate those factors before explaining why the result changed.
A practical test table should include original format, output format, dimensions, byte size, visual acceptance and the exact detection response. Keep unavailable checks separate from negative ones.
Choose a format for the delivery requirement
Use the destination’s needs to guide the choice: transparency, compatibility, exact samples, download size or further editing. Keep the original master separately. Avoid repeated conversions between lossy formats when a single export from the master will do.
There is no reason to cycle through formats until a file becomes hard to inspect. A small file that destroys a label or shifts a product color can be a poor result regardless of its detector score.
For delivery choices, read the PNG, JPEG and WebP comparison. For watermark questions, use the verification workflow. This article explains format behavior; it does not report a new cross-format SynthID benchmark.