Seedream Detector

Add a Seedream picture to the checker above and your browser reads its Content Credentials first, without uploading anything. Every Seedream original we made carried a signed record saying it was AI-generated, but the signer is not on the trust lists we check, so the Seedream detector treats that record as a claim and asks a detection model to judge the pixels as well. In our test it got all 7 Seedream images right.

Nothing extra is needed from you. If the credentials and the model agree, the answer is Likely AI-generated, with the signer named in the reasons. If they disagreed, you would get Can't tell rather than a verdict built on a label alone. This is the same flow as our main AI image detector, with the numbers below measured on Seedream files only.

Yes. All 5 Seedream originals we generated ourselves, with Seedream 5.0 Pro and Seedream 4.5, came back with C2PA Content Credentials that say the picture was made by AI. The record is signed by Byteplus Pte. Ltd., part of ByteDance, and the software it names is c2pa-rs, the open-source C2PA library that writes the record, rather than Seedream itself.

That signature is real, but it doesn't settle the question on its own. Content Credentials are trusted when the certificate that signed them is on a recognised trust list. BytePlus's certificate was not on the current C2PA trust list or the older interim list that our checker uses, so the checker can confirm what the file says but not who said it. You can read the full record yourself with the Content Credentials checker.

Because anyone can write an AI label into a file, and a signature only proves something if the signer can be checked. Our rule is the same for every generator: credentials from a signer on the trust lists settle it, and everything else, including signed records from unlisted signers and plain metadata labels, goes to a detection model.

For Seedream that means 7 of 7 images went to Sightengine, 5 of them carrying the BytePlus record. The model agreed with the record every time. The extra step costs one of your free checks, but it protects against the opposite case: a real photo with an AI label copied into it, which a label-only checker would call AI. Two of the six families in our test, FLUX and Seedream, are in this position; the FLUX detector page shows the same rule at work.

Sightengine, the main engine behind this Seedream detector, scored all 7 Seedream images as AI, and none came close to the line: the lowest score was 0.99, well above the 0.8 our checker needs before it says Likely AI-generated. Hive, our backup engine, caught 6 of 7 at its recommended 0.9 threshold.

Seven images is a small sample, so read these as a good sign for the two models we tried, not as a measured error rate. Across all six generator families, the accuracy report has the full method, the per-family results and the limits of the test.

A screenshot of a Seedream image loses the credentials, just like any other screenshot: 0 of our 2 Seedream screenshots kept them. Your phone draws a new image from the screen, and the signed record belongs to the original file only. Many social apps strip it on upload too.

The pixels survive, though. Sightengine still scored both Seedream screenshots above the checker's 0.8 line (2/2). If you are checking a Seedream image you found in a feed, paste the screenshot with Ctrl+V and expect a slightly wider range than you would get on the original. Our screenshot guide explains what else to try, such as finding the first post or a larger copy.

It can't tell you which model or version made a picture, only whether the picture is likely AI-generated and why. A missing BytePlus record proves nothing either: screenshots, re-uploads and edits remove it, and apps built on Seedream may write different metadata or none at all.

We tested only pictures we generated through an API, with Seedream 5.0 Pro and Seedream 4.5, so edited outputs, collages and pictures from older Seedream versions are outside what we measured. As with any check, treat Likely AI-generated as a reason to ask questions, and use the visual checklist when you can see the whole scene.

No. It can say a picture is likely AI-generated and show the BytePlus-signed record if the file still has it, but it doesn't identify the model. A signed record from a signer on the trust lists would be close to proof; BytePlus's signer isn't on the lists we check.

Because the credentials alone don't settle it, so the picture goes to a detection model. If someone already checked the exact same file, the stored result comes back without a new check. You get 3 free checks a day.

None of the 5 original files we generated had one we could see, corners included; the AI disclosure we found was in the Content Credentials. Apps that use Seedream may add their own marks, which a crop can remove.

Yes. Screenshots have no credentials, so the model decides. Sightengine still caught 2 of the 2 Seedream screenshots in our test.