AI Content Policy

Scope and Purpose

FORMÉ is a business-to-business platform that generates fashion and product media — product shots, styled compositions, and short video clips — for brands, retailers, and the agencies that work for them. Generation is driven by prompts and configuration supplied by the account holder.

This policy is written to be medium-independent. Everything below governs any output the Platform generates, whether a still image, a video clip, or any audio the Platform may generate in future. Where a rule or a safeguard applies differently to moving image than to still, that difference is stated explicitly rather than left to inference.

This policy states, in plain language, what the platform's AI systems may and may not be asked to produce, and how we enforce those limits. It applies to every person who can prompt a model through FORMÉ: account holders, each seat they invite, and any third-party vendor, agency, or contractor operating the account on their behalf. Account holders are responsible for the conduct of everyone they grant access to.

This page is a summary written to be read. The binding rules are in our Terms of Service and Privacy Policy; each section below links to the clauses that govern it. Where this summary and those documents differ, those documents control.

1. People in Generated Outputs

Fashion media involves human figures, so this is the area we govern most tightly. A human figure can reach an output on exactly three paths, and each has its own gate. The gates apply to video exactly as they apply to a still: a clip is an output like any other, and nothing about setting an image in motion relaxes the consent, likeness, or age rules below.

1.1 Virtual models are synthetic and adults-only

Virtual models are AI-generated figures that are not based on, or derived from, any specific real person's likeness (Terms §4A.4). They may only depict adults. The age selector offers no value below 18, and free-text prompts describing a person are screened for references to minors and rejected before any generation request is made. That screen is a safeguard layered on top of the adults-only design, not a substitute for it — prompting for a minor is a policy violation whether or not an automated check catches it.

1.2 Real models appear only with recorded consent

Where imagery features a real person, that person is a model contracted through the platform, and consent is a sequence of recorded states rather than a checkbox. A brand submits a hiring invitation describing the intended campaign and usage. The model reviews it and may decline. Only after acceptance can outputs be generated, and those outputs are proposals with no licence attached (Terms §4.4). The model then reviews the generated images individually and approves or rejects each one. Only approved images are released and licensed (Terms §4). A model is never obliged to release, and consent is scoped to the declared usage — changing the campaign, the territory, or the media requires a new approval (Terms §5.3).

Biometric identifiers and appearance data collected to build a model's digital twin are governed separately, including consent, retention limits, and destruction (Privacy §4).

1.3 Uploaded photographs require rights you actually hold

Uploading a photograph of a person is an affirmation that you hold the rights to use it for AI generation and have the consent of the person depicted. Uploading imagery of someone who has not consented — including a public figure, a celebrity, or a private individual — is prohibited, as is uploading content you do not have the rights to (Terms §10.14).

1.4 Video inherits the gate of whatever it depicts

The Platform generates short video clips. A clip may be conditioned on a virtual model or on a contracted model's approved imagery so that the figure stays recognisable across frames. Because the figure is the same figure, the consent position is the same: a clip depicting a contracted model is a proposed output requiring that model's approval before release, and its licence is bound to the same declared usage as a still. Generating a clip is not a way to obtain a depiction that was not approved as an image.

Two limits are specific to moving image. Motion must not be used to place a person in conduct, speech, or a setting they did not approve, which is a form of the impersonation and false-endorsement prohibitions below rather than an exception to them. And video clips are generated by Google's video model, which embeds SynthID, Google's invisible watermark, at generation. Removing, suppressing, or degrading it is prohibited (Terms §10.10).

1.5 Voice

The Platform does not currently generate audio, speech, or synthetic voice, and the video clips it produces are silent. We state the rule in advance rather than after the fact: a person's voice is part of their likeness. If the Platform generates voice in future, it will sit behind the same consent gate as their appearance — recorded, scoped to a declared usage, and revocable — and the prohibitions in section 2 apply to audio outputs on the same terms as visual ones. Using Platform outputs as source material for voice synthesis elsewhere is already prohibited (Terms §10.8).

2. What the Model May Not Be Asked to Produce

The following are prohibited on FORMÉ, in any medium — still image, video, or audio. This restates Terms §10, which governs in full.

• Sexual or exploitative content. Pornographic, sexually explicit, obscene, or sexually abusive depictions of anyone (§10.2). Any depiction of a minor, in any context, is categorically prohibited.
• Impersonation and synthetic likeness of real people. Using a virtual model or any output in a way designed to impersonate, or be confused with, a specific real person — including face-swapping a real person onto generated media, and including any clip that depicts a real person moving, acting, or appearing to speak without their consent (§10.4).
• False endorsement. Implying that a model personally used, endorsed, sponsored, or is affiliated with a product, company, cause, or message beyond the approved campaign scope (§10.4).
• Harassment, defamation, and abuse. Outputs used to defame, harass, threaten, humiliate, stalk, or intimidate any person, including the model depicted (§10.3).
• Fraud and deception. Scams, false advertising, or misrepresentation of a product's material characteristics (§10.5).
• Political and sensitive-category use without approval. Political, electoral, religious, or comparable sensitive messaging, which requires separate approval (§10.6).
• Unlawful and harmful content. Anything illegal in the jurisdiction of use, or that promotes violence, self-harm, or hatred against a protected group (§10.1).
• Tampering with provenance. Removing, falsifying, or suppressing watermarks, metadata, or other authenticity and rights-management markers (§10.10).
• Retraining and extraction. Using outputs to train, fine-tune, or develop another AI model, and scraping or bulk-extracting platform assets (§10.8, §10.11).
• Editing that changes meaning. Colour grading and crops for production are permitted; altering a model's body, face, or identity, or placing an approved output into a context the model did not approve, is not (§8).

3. How These Limits Are Enforced

Enforcement is layered, because no single layer is sufficient. Some layers are ours and apply whichever model runs a request. Others belong to the company that provides the model. Google runs every request by default. A user may choose another provider's model for some image requests; the providers we use are listed in Privacy §5.2. Product uploads and product processing always use Google's models, and so does all video. A Studio shot, including a shot of products alone, may use another provider's model that a user chooses; that provider must not train its AI models on what we send, and must hold it under confidentiality terms.

• Before the request (ours). Structured inputs constrain what can be expressed at all — the age selector has no sub-18 value — and free-text describing a person is screened for minor references and refused before it reaches a model. A photograph uploaded to create or enhance a model portrait is checked automatically for apparent minors and explicit content, and refused if the check finds either or cannot run.
• Consent and likeness (ours). A contracted model's likeness is used only within a booking they accepted. A request that may show a real person — a contracted model, or anyone in an uploaded photograph — goes only to Google's models or to a provider we have approved for real people. That provider must not train its AI models on what we send, and must hold it under confidentiality terms. Any other request of this kind is refused before it is sent.
• At the model provider (theirs). Each provider applies its own safety systems and may refuse a request or withhold an output. For Google's image models, we also set Google's safety settings to block sexually explicit, hateful, harassing, and dangerous content at the medium threshold and above. Users cannot change them. Google's video model does not accept these settings, so video relies on Google's built-in filters and on our checks before the request. Other providers apply their own policies, which we do not configure.
• After generation (ours). We check every image from a provider other than Google with our own safety classifier before you see it. It looks for apparent minors, sexual content, violence or gore, and hate symbols. An image it flags is withheld and its credits are refunded, and so is an image the check could not run on: nothing from these providers is shown unchecked. Like any automated check, it can miss content or flag content wrongly. Images from Google's models are covered by the Google safety settings above and are not checked again.
• Before release. Where a real model is depicted, that model approves or rejects each generated output individually, clips included. Unapproved outputs are never licensed and cannot be exported. Automatic checks also review many studio outputs for accuracy, such as product colour, shape, and label text. These are quality checks, not a content-safety screen.
• Provenance. Images and video from Google's models carry SynthID, Google's invisible watermark. Google's images also carry C2PA Content Credentials. Other providers may add their own provenance data, such as C2PA Content Credentials. For an image from a provider other than Google, we keep the provider's original file, including any provenance data it carries, except images of a booked model, which follow the booking's licence terms. We show and deliver images as PNG or JPEG, so when a provider returns another format, the image you see is a converted copy; downloading it from your Library gives you the provider's original file. We do not remove a provider's provenance data on purpose. But some steps create a new image file that may not carry it, for example a format conversion, a preview, an edit, or a file sized for a sales or social channel. Users must not remove or alter provenance data (Terms §10.10).
• At the account. We may review content, refuse or remove it, require remediation, suspend generation, and terminate accounts. We may act immediately where there is risk of harm (Terms §13).

We reserve these review rights but do not undertake to monitor all activity (Terms §13.6). Compliance is a shared obligation, and the account holder is accountable for what their account generates.

4. Your Content and Model Training

We use third-party AI infrastructure to operate the platform. Your prompts, uploaded content, and the resulting outputs are processed by those providers only as needed to deliver the service. We use a model provider only if it commits not to train its AI models on the content we send. We do not authorise any provider to use your content — including personal information, appearance data, biometric identifiers, garment content, or outputs — for its own purposes. Some providers keep content for a limited time under their own terms. The providers, and what we know of how long each keeps content, are listed in Privacy §5.2.

We may use aggregated, de-identified quality metrics to improve our own platform systems. That does not involve sharing identifiable content. What our AI processing does not do is set out at Privacy §5.3.

5. Reporting a Violation

If you believe content generated or published through FORMÉ violates this policy — including if you are a person depicted in imagery you did not consent to — contact us at support@forme.style. Include enough detail to identify the content. We investigate reports and may remove content, require remediation, or suspend the responsible account (Terms §13.4).

Models contracted through the platform retain their rights in their own likeness and may raise concerns through the same channel (Terms §9.3).

6. Changes to This Policy

We update this policy as the platform's capabilities change, and the "Last Updated" date above reflects the most recent revision. Material changes to the binding rules are made through the Terms of Service, which sets out how changes take effect.