Documentation/Practical guides

Practical guides

AI labeling requirements for generative image and video apps

If your app generates images or video, start by checking the requirements for your product role and markets. Ariadne Mark provides a free, open-source implementation kit: watermarking, C2PA provenance, a verifier and a recovery registry. It runs on your own Linux CPU backend.

Evaluating whether it fits your product? Check the stack, runtime, costs and production steps in one place. For a first technical result, use the release quickstart.

Does my app need to label AI-generated images?

The practical question is which duties apply to your system, output and distribution. A small team, free app or entertainment use does not by itself settle that assessment. Use the EU AI Act guide and California AI Transparency Act guide to check the relevant role, dates and exceptions.

For an applicable visual-media workflow, distinguish machine-readable marking, user-facing disclosure and a usable verification route. They serve different purposes. The requirements matrix connects each relevant measure to the toolkit and remaining operator actions.

What applies now, if relevant, and later?

This navigation is dated 15 November 2026. Start with your role, markets and the system's first placement date; team size alone is not an exemption. The linked jurisdiction guides contain the detailed scope and exceptions.

Situation When to assess it Your next action
You provide an image/video generation system in the EU Now: Article 50(2) generally applies from 2 August 2026 Implement machine-readable marking and detectability; record your role and technical evidence
That EU system was placed on the market before 2 August 2026 Provider transition ends 2 December 2026 under Article 111(4) Record the actual placement date; do not apply this transition to all deployer duties
Your product or users deploy realistic synthetic people or events If the output qualifies as a deepfake Assess clear user-facing disclosure separately; creative/satirical use affects its form
You produce a GenAI system publicly accessible in California Assess the amended provider rules now Do not use the former one-million-monthly-user threshold; check latent disclosure, free verification and operating duties
You run a website offering model source code or weights for download, or a large content platform If you meet that separate actor definition Review the separate platform provisions and their 2027 dates; self-hosting your own model is not itself this hosting-platform role
You also generate synthetic speech or music Before offering that modality This visual kit does not independently mark generated audio

The EU source is Article 50, with the pre-existing-system transition in Article 111(4). California navigation follows the chaptered SB 1000 and the dated source review. Old search snippets may still describe an earlier California threshold. There is no single “US AI Act” covering every state or use case.

For a small app, take four concrete steps: map the role and final media; choose and test a marking profile; deploy the verification route; complete product disclosures and production signing. Use the matrix to keep technical measures, operator duties and voluntary Code measures separate.

Two common app scenarios

Your product Where labeling fits Start here
A mobile app generates greeting cards, memes or short clips on your own servers The backend marks each final rendition before publishing or returning the download Python integration
A web app uses a generation API that already returns Content Credentials Inspect and preserve the existing provenance; assess the downstream role and supported parent-history workflow before adding another export Upstream-input boundary

The kit is a backend component. It is not an on-device iOS/Android SDK. A Node.js or Next.js product can use a Python worker or the CLI from its own background queue; see the Node and mobile-backend guide.

Can I use this with Stable Diffusion, FLUX or ComfyUI?

Integrate at the exported-file boundary on your own backend: pass the final supported PNG, JPEG or SDR MP4 to the worker. The SDK does not need to change model weights or the generation API. Evaluate the files your particular model and delivery pipeline produce.

There is no packaged ComfyUI node in this release. Use a file/job hand-off to the Linux worker, or build and test a dedicated adapter. The generic backend pattern is documented; a named-model plugin is not claimed as qualified.

From a requirement to a working implementation

  1. Identify the role and markets. Record who provides the system, what users can generate and where the output is offered. Use source-linked jurisdiction guides for the actual assessment.
  2. Prepare the final file. Crop, resize, add overlays and choose the final encode before marking. Each published rendition needs its own workflow.
  3. Mark and sign. Add the recoverable watermark ID and signed provenance with truthful public provider, system and creation information.
  4. Make verification available. Deploy the verifier and persistent registry where your product relies on them. Give recipients a clear way to reach it.
  5. Apply any required visible disclosure. An invisible watermark does not automatically replace a human-readable label. Compare the layers.
  6. Test the delivered file. Check bytes downloaded through your own app/CDN, including your normal transformations.
  7. Keep the evidence. Save the receipt, configuration and relevant test results with the release. Use the operator checklist.

What can I install for free?

The release quickstart provides a first-export route for the qualified Linux x86-64 / Python 3.12 runtime. It produces a marked file and a verification result using development credentials. The deployment kit connects SDK, verifier and registry.

There is no Ariadne per-export API fee. Your own compute, storage, hosting and production signing or timestamp services remain operating costs. See the measured CPU costs before sizing your worker.

How should I choose a video profile?

Economy5s and Economy2s reduce watermark inference. Resilient computes a fresh signal for every frame. Use the current CPU measurements together with the profile limitations, especially for short clips and excerpts. The profile name is a technical setting, not a legal approval.

Before enabling the workflow for users

Replace evaluation credentials with your actual production identity and trust configuration. Confirm applicability, visible disclosure where required, verifier availability, data handling and delivery behavior for your product. Ariadne supplies implementation tools and evidence; the checklist remains tied to your app's facts.

Keep three milestones separate: a development export that verifies, a pilot integrated into your own app, and an operational production deployment. The first does not establish the other two. The toolkit includes the authenticated TSA integration;your actual production CA/TSA credentials and live export still need onboarding and verification; the current timestamp support boundary is an implementation dependency, not just a form to fill in.

Choose a video mode · Get your first marked export · Try the real hero image · Inspect downloadable SDK samples