Generated images are easy to make and awkward to publish, because the rules are spread across a model's terms, a provider's tier structure, and the law of wherever you and your audience are. This page is a checklist, not legal advice, and it exists because we kept being asked the same four questions.
The short answer
Answer four things before an image goes live: does the provider's terms allow commercial use on your tier, does it indemnify you or merely license the output, does it contain a plausible likeness of a real person, and does it imitate a protected brand or a living artist's style. If any answer is unclear, do not publish.
1. Commercial use and the tier you are on
Most providers allow commercial use, and several distinguish free-tier output from paid output. The distinction is usually in the terms rather than the interface, which is why teams get it wrong.
Check the exact plan you generated under. An image made on a personal free account and used in a client campaign is the most common version of this mistake.
2. Licence versus indemnification
These are different promises and they get conflated constantly.
| Term | What it means |
|---|---|
| Licence | You may use the output |
| Indemnification | The provider defends you if a third party claims the output infringes |
Adobe Firefly is positioned around the second. Most others offer the first. If your client contract expects indemnification, that single row may decide your tool choice regardless of image quality.
3. Likeness
A generated face is not automatically fictional. If the output plausibly resembles a real person — a celebrity, a public figure, someone who could be identified — publication carries risk that model terms do not cover.
Two practical tests: would a stranger name a real person on seeing this image, and would you be comfortable if that person saw the campaign. If either answer is uncomfortable, change the prompt.
4. Style imitation
Prompts that name a living artist or a specific brand's visual identity are the area where terms and law are least settled. "In the style of [living artist]" is a request that a provider may allow and a rights holder may still dispute.
The safer version is descriptive rather than attributional: "high-contrast black-and-white portrait with hard shadows" rather than a named photographer.
What about open-weight models?
FLUX publishes open weights for some models, which changes the analysis. The licence attached to the weights governs what you may do with the model, and the provider of the *serving* stack governs the output. Self-hosting gives you control, not immunity.
It also shifts responsibility: when you run the model, there is no provider in the middle to indemnify you.
A publishing checklist
- Which account and tier generated this image?
- Does that tier permit commercial use in this context?
- Is indemnification required by the client, and does the provider offer it?
- Does the image contain a plausible likeness of a real person?
- Does it reproduce a logo, trademark, or recognisable brand asset?
- Does the prompt name a living artist or a protected character?
- Is the model's version recorded so the image can be regenerated identically?
The last item is operational rather than legal, and it is the one that saves you when a client asks for a variation six months later.
The contract question
Most licensing trouble is not about the model; it is about the contract the image ends up in. Client agreements increasingly specify that delivered assets are owned outright, or that the agency indemnifies for intellectual property claims.
If the contract says owned, check what the model provider's terms actually grant — a licence is not ownership. If it says indemnified, check whether the provider offers indemnification at all, and on which plan. These are the two clauses that determine which tool is usable for client work, and they are the ones rarely read.
Recordkeeping that saves you later
Keep three fields per published image: the model and version, the prompt, and the date. That record answers the questions that arrive six months later — "can we make a variant", "what generated this", "when did we switch models".
It is also cheap. A spreadsheet is enough, and the alternative is regenerating an image you cannot reproduce.
Frequently asked questions
Do I need to disclose that an image is AI-generated? Some platforms require it and some industries expect it. Disclosure also protects you if viewers notice artefacts.
Can I copyright a generated image? Rules vary by jurisdiction and are unsettled for purely machine-generated work. Human modification strengthens the position, which is one more argument for editing rather than publishing raw output.
What if the model output resembles an existing stock photo? Rare, and hard to prove either way. Recording the prompt and model version is the practical mitigation.
Last checked: 9 November 2026. Terms change; re-read them before a campaign, not after.
Related: FLUX vs Firefly vs Recraft, design tools vs design models, and image tools.