Nano Banana watermark: SynthID and C2PA on images via Sume

Google says Nano Banana 2 images carry a SynthID watermark and C2PA Content Credentials, and that Nano Banana 2 Lite images carry SynthID. Sume's image docs make no statement about either, and they say results come back as Sume-hosted, signed URLs, so whether a mark survives the trip is something to test on your own downloaded file.

Google's side is from its two launch posts, Sume's from the Image API docs, read 2026-09-29.

What does Google say is added to Nano Banana images?

The two posts differ, so keep them apart.

What does Sume return?

The Image API returns data[].url, a Sume-hosted signed URL, rather than inline base64, because Sume mirrors generated media. Sume's docs list this as the result payload; they do not describe the file's metadata or say whether it keeps or drops a provider's marks.

So this post cannot tell you that a Nano Banana image from Sume still has SynthID or C2PA data. It also does not say it lacks them.

How do I check a file I downloaded?

Generate one image, download it, and inspect that file with the verification tools Google or the C2PA community publish. Repeat when you change models, since the two Google posts describe different marks. Treat a positive result as evidence for that file only.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "google/nano-banana-2", "prompt": "a red bicycle against a white wall"}'

Should I label AI images myself?

That depends on the platform and the law where you publish, and neither is settled by a watermark. Do not rely on an embedded mark alone; follow the disclosure rules of the place you post.

Why might a mark not survive?

I have no evidence either way, so this is a list of things to test rather than a claim. Marks can be lost when a file is re-encoded, resized, cropped or screenshotted, and Sume's docs do not describe how its mirroring stores the file. That is why the check belongs on the exact file your users will see, after any edits your pipeline makes.

Does the answer change by model?

It can. Google describes SynthID and C2PA for Nano Banana 2 and SynthID for Lite, and the pages read say nothing about Nano Banana Pro, which Sume also lists as google/nano-banana-pro. Test each model id you ship rather than carrying one result across.

What should I record for my own files?

Keep the model id, the request, the job id and the date next to each image you publish, so you can show where it came from. Store the original download unedited alongside any edited copy. If provenance matters to your customers or your platform, test the original file with a verification tool first, then the edited copy, and write down which one passed.