Image output carries a visible AI-generated label. Video output depends on the model.

We would rather you find this out on this page than after you have written the integration.

What we tested

One identical prompt per model, output downloaded and inspected directly — visually and with OCR on corner crops. Images tested 2 August 2026; full re-test of all nine media models 4 August 2026 (video: first, middle and last frames).

Three honest caveats

• No visible label does not mean no labelling at all. Providers may still embed invisible metadata watermarks — CogVideoX’s cover image, for example, is served through Zhipu’s watermark pipeline even though nothing is visible. We only report what we can see and OCR.
• This is upstream behaviour, not ours. Providers can change how they apply labels at any time. If our next scheduled check finds a change, this page will be updated.
• Paying never removes an image label — all five labelled models above except cogview-3-flash are paid. The pattern is provider × media type, not free vs paid.

Why the labels are there — and why they differ

China’s Measures for the Labelling of AI-Generated Synthetic Content, in force since 1 September 2025, require providers to attach labels to generated images, audio and video. Every model above is operated inside mainland China. How each provider implements that obligation, however, is up to them: some burn a visible corner label, some embed only invisible metadata, and behaviour can differ between their image and video pipelines. The table above is what we actually observed through our API on the test dates.

Two consequences worth stating plainly:

• It is not a NovAI setting. We proxy the provider’s response. There is no parameter we can pass to disable a label, and adding one would mean asking the provider to break its own compliance obligation.
• We will not remove labels. The same rules prohibit deleting or altering them. We are not going to build a watermark stripper, and we would advise you not to either.

What this rules in and out

Image models (all five): plan around the label

• Good for prototyping, internal tools, testing prompt pipelines at zero or near-zero cost, clearly marked AI demos, research output, and quality/latency comparisons.
• Not suited to client deliverables, stock assets, ad creative or anything shipped as clean artwork — the label reads as Chinese text in the corner.

Video models: check the table per model

• doubao-seedance-2.0, doubao-seedance-2.0-fast and cogvideox-flash returned visually clean output in our tests. If your use case permits visible-label-free video, these are currently the options — with the caveats above (possible invisible metadata; upstream behaviour can change).
• hy-video-1.5 carries the corner label, so the same restrictions as the image models apply.

Questions we expect

Is Seedance video really clean?

In our 4 August 2026 test (first, middle and last frames, visually plus OCR on corner crops at 3–4× zoom), yes — no visible label. We cannot rule out invisible metadata, and ByteDance can change this upstream. Re-check this page if you are building something sensitive to it.

Can I crop the label out of images?

Technically the label sits in a corner, so yes, you could. We are not going to help you do it and we do not recommend it. Beyond the compliance problem, cropping changes the output dimensions you asked for.

Will you add an unlabelled model?

Only if we can source one from a provider that has no such obligation. We have not done that yet, and we are not going to announce it before it works. If and when it happens, it will be listed on this page.

Does this affect the free tier claim?

No. cogview-3-flash and cogvideox-flash really do cost $0 with no credit card. A label — where present — is a limitation of what you get, not a hidden charge.