A single reference asset is a data point. A brand system is a rule set, and the vendors building the tools now say so directly.
A creative director finds one campaign visual that finally looks right. The palette lands, the composition works, the tone matches the brand. The next brief goes out with that image pasted in as make it look like this. Three weeks later, five variations come back that each resemble the reference differently, because nobody wrote down what made it work.
The Incumbent Way: Rebrief From the Reference Every Time
Most marketing teams treat a strong piece of creative as a mood board, not a specification. Every new request starts the same way: attach the reference, describe the vibe, hope the output lands close enough. Nobody separates what has to stay fixed from what is allowed to change, so every generation becomes a fresh negotiation instead of an application of a known rule.
Adobe's own product guidance for training Firefly Custom Models states the requirement in numbers: choose "at least 10-30 high-quality images that showcase the brand-specific styles and concept subjects you want to achieve," a varied set shot from "various viewpoints and backgrounds while maintaining a consistent aesthetic," tagged with a minimum of three Model Tags describing only "permanent attributes," with changeable attributes explicitly excluded (Adobe Firefly best practices). That threshold is specific to Firefly Custom Models. It is not an industry standard and no vendor claims it is. What it proves is narrower and more useful: even the company selling the generation tool will not train a brand model on one image, because one image cannot show a system what stays fixed and what is allowed to move.
The Cost of Skipping the System
Skip the extraction step once and the cost is a round of revisions. Skip it across a campaign and the cost compounds into brand drift nobody can trace, because there was never a documented rule to violate. Skip it at the scale a large marketing org now generates content, with agents drafting decks and campaign copy on demand, and the drift stops being visible one asset at a time. It shows up as a brand that no longer looks like itself, discovered only after the damage has shipped.
What Production Scale Actually Looks Like
Coca-Cola supplies the current evidence at real scale. On July 20, 2026, the company launched what it calls "a new global visual identity system" across "more than 200 markets," rolling out through Latin America, Europe, the Middle East, and Asia, backed by "a new immersive Brand Center alongside tech-powered Design Intelligence tools" built to "support governance, craft, and consistency at scale" for internal teams and for "the wider network of agency partners" (Coca-Cola, July 20, 2026). The release names four deliberately narrow invariants carried through every market: the red and white palette, the Dynamic Ribbon, the Arden Square, and the Spencerian script. That is a fixed layer stated in public, at the scale of a top-five global brand.
Coca-Cola's earlier announcement, from May 2025, described a separate initiative called Project Fizzion, which "captures creative intent and encodes it into a StyleID, a machine-readable identity that automatically applies brand rules across formats, platforms, and markets" (Coca-Cola, May 15, 2025). That architecture, rules extracted once and applied automatically at generation time, is the clearest public description of a machine-readable brand system. The July 2026 rollout does not name Fizzion, and nothing in the July release states that the global identity system runs on it. Treat the two as related evidence of the same direction, not the same product.
Independent vendors describe the identical structure without any connection to Coca-Cola. Adobe's Firefly Custom Models API splits brand encoding into Subject Models, which represent "specific characters, products, or objects," and Style Models, which capture "color palettes, patterns, brush stroke techniques, or illustrative cues" so that "all generated visuals share a consistent, recognizable style" (Adobe Firefly Services docs). Figma builds the same separation into Figma Make, where teams "use your Figma style libraries and code components as the foundation for explorations" so that "every prototype your team creates in Make stays on-brand and system-aligned" (Figma). Salesforce stores brand identity and tone as structured record fields that an AI agent reads at generation time, and it advises customers to "establish an approval process for creating, updating, and publishing brand styles" because by default "all contributors in a Marketing workspace can create or edit your company's brand styles" (Salesforce Help). Three vendors with no relationship to each other, and no relationship to Coca-Cola, converge on one architecture: invariant rules extracted once, variable concepts generated inside them.
The System Is Governed, Not Frozen
A brand system that never updates is a liability with better production values than a PDF. Adobe frames its Brand Intelligence work as moving organizations "beyond static artifacts, brand guidelines, into a system that learns from qualitative and nuanced inputs such as review cycle feedback, annotations, rejections and approvals" (Adobe). Every rejection a reviewer issues, every approval a reviewer grants, becomes an input the system uses to hold its line more precisely next time. Salesforce documents the version-control cost of skipping this discipline: when a default brand record is updated, "the update isn't applied to the published content," and someone has to unpublish, remove the old brand, and reapply the new one by hand (Salesforce Help). A governed system versions its rules and tracks what ran under which version. A frozen system just gets stale until someone notices.
The Fork
You have two ways to scale a strong piece of creative. Paste it into every new brief as a vague reference and get back interpretations that each resemble it differently. Or extract its invariant rules once, version them, and generate variable concepts inside that structure on every future brief, feeding each approval and rejection back into the rules themselves.
That is not a creative process preference. That is the difference between one good asset and a brand that reproduces its own quality on demand.
Verdict
A reference image tells you what worked once. A system tells you why, and lets you get it again, on the two-hundredth market as easily as the first. The brands treating every new asset as a fresh negotiation with a mood board are the ones drifting one approval at a time. The ones extracting and versioning the invariant rules are the ones whose two-hundredth market still looks like the first.
Write down what must never change. Version it. Let everything else move.
Work With Magnet
Magnet builds brand architecture and creative systems that turn one strong reference into a governed, repeatable standard. See how Magnet approaches brand architecture at https://www.magnet.co.
Sources
- Adobe, Firefly best practices for training custom models: https://helpx.adobe.com/firefly/web/work-with-enterprise-features/train-custom-models/best-practices-to-effectively-prepare-and-create-custom-models.html
- Adobe, Firefly Services API docs, Custom Models Overview: https://developer.adobe.com/firefly-services/docs/firefly-api/guides/concepts/custom-models/
- Adobe, Brand Intelligence announcement: https://news.adobe.com/news/2026/04/adobe-introduces-brand-intelligence
- Coca-Cola, global visual identity system launch, July 20, 2026: https://www.coca-colacompany.com/media-center/coca-cola-sharpens-its-identity-under-one-bold-visual-system
- Coca-Cola, Project Fizzion introduction, May 15, 2025: https://www.coca-colacompany.com/media-center/the-coca-cola-company-introduces-fizzion
- Figma, AI design systems generator in Figma Make: https://www.figma.com/solutions/ai-design-systems-generator/
- Salesforce Help, brand styles documentation: https://help.salesforce.com/s/articleView?id=mktg.mktg_content_brand.htm&language=en_US&type=5
- Magnet: https://www.magnet.co
