Augment the pipeline. Never fabricate identity.
Every face in a fashion campaign is either a real, licensed person or a licensed digital twin of one. No fully synthetic principals. No prompts against another brand's campaign, another photographer's style, or a named person's likeness. AI raises throughput — it should not raise identity risk.
This is the one rule the rest of the playbook enforces. Every section that follows is either a way to produce imagery inside the rule, or a gate that stops imagery that breaks it.
Three ways AI enters production.
- Digital twins of licensed models. Written twin licence per model — use cases, term, territory, compensation, rev-share. EU-based processing, GDPR-compliant vendor. This is the preferred path for talent-led campaigns. Real people consent to their AI likeness, are paid for it, and share in the value it generates.
- AI-assisted post-production on real photography. Retouching, background variants, wardrobe swaps applied to genuine shoots. Real talent, model release updated to expressly cover AI processing, style transfer, and downstream reuse.
- Product-only generative work. Invisible-mannequin, flatlay, and on-body renders using licensed twins. No principal talent present in-frame. Highest-throughput path, lowest-risk profile.
What you do not produce.
- Fully synthetic "new person" faces used as the campaign's principal talent.
- Prompts referencing a named person, competitor brand, specific past campaign, or a named photographer's aesthetic.
- Image-to-image or reference-image workflows fed with third-party campaign photos, editorial images, or scraped material.
- Any AI processing of a real model's image without an updated release covering AI use, style transfer, and downstream reuse.
Non-negotiables for any AI imaging partner.
- Written reps that all training data is licensed.
- Indemnity for right-of-publicity, IP, and moral-rights claims — floor of SEK 10M / €900k / $1M, or the vendor cap, whichever higher.
- Right to audit training data on 30 days' notice.
- Output kill-switch if the vendor's base model loses an IP suit.
- EU data residency for any biometric or digital-twin processing.
Every asset clears seven gates before it circulates.
Interactive checklist — tick as you clear each gate for the current asset. State persists per Asset ID in your browser. Use it live, or print it as a QA sheet.
Retain seven years, per asset.
Prompt text · model version · seed · loaded adapters · timestamp · operator · vendor invoice reference · twin licence reference. Kept in the DAM alongside the final asset, indexed by Asset ID. No exceptions. When a claim lands three years from now — and one will — the audit file is what stops it in ten minutes instead of ten months.
One line on published creative.
"Imagery includes AI-assisted production." Covers EU AI Act Art. 50, the New York synthetic-performer law effective 9 June 2026, and California and Tennessee right-of-publicity statutes. For paid placement on Meta and TikTok, also self-declare AI content in each platform's ad manager at upload. One default line, applied globally, saves you from writing per-jurisdiction rules that will drift out of date within twelve months.
Update before the next shoot.
- Model releasesAdd AI-processing, style-transfer, digital-twin, and downstream-reuse clauses. Retro-fit key existing talent.
- Agency MSAFlow down every vendor term in §04. Right-to-audit passes through to any sub-processor.
- InsuranceAdd a media-liability rider covering AI-generated content. Roughly SEK 100–250k / €9–22k per year.