Build an On-Brand Asset Production Pipeline

Workflow area: Design Ops · Tier: Foundation · Hands-on time: ~30 minutes to stand up

The volume problem is real: empty states, onboarding art, blog headers, seasonal variants, localized imagery — a mid-sized product consumes hundreds of assets a year, and "we'll commission it" quietly becomes "we'll skip it." Generation solves volume. What it doesn't solve is coherence and rights: fifty assets that each look plausible and collectively look like fifty different products, generated under terms nobody read. The pipeline is style codification plus art direction plus policy — the generation step is the easy 20%.

When to use — and when not to

• Use it for volume asset classesempty states, spot illustrations, background textures, marketing variants; anywhere the style repeats and the subjects vary.
• Use it for exploration and compstrying five illustration directions before commissioning one is the cheapest brand decision you'll make this year.
• Use it for variant fan-out of approved artsizes, crops, seasonal recolors and localized contexts of an already-directed asset.
• Skip it for the brand's identity corelogo, mascot design, the hero illustration that IS the brand: commissioned, owned, human-made, with contracts that say so.
• Skip generated humans where authenticity is the promisetestimonial imagery, team pages, "real customer" contexts; a synthetic person in an authenticity slot is a trust violation waiting for a screenshot.

The recipe

1. 1.

   Codify your visual style as a written spec. Gather 6–10 existing approved assets and have a vision model reverse-engineer the style in words:

   Copy

   These are approved brand illustrations. Write a STYLE SPEC that
   would let an image model reproduce the style without seeing them:
   - Line quality, weight, corner treatment
   - Color palette (name hexes if visible) and how colors are
     distributed
   - Composition rules: density, negative space, perspective
   - Human figures: proportions, diversity representation, level of
     abstraction
   - What this style NEVER does (gradients? photorealism? outlines?)
   [ATTACH 6-10 APPROVED ASSETS]
   

   Edit the spec against your actual brand guidelines — the model describes what it sees, and what it sees includes accidents.

2. 2.

   Build the master prompt and lock it as a saved style. Combine the spec into a reusable prompt block. In Midjourney, encode it as a saved style/profile reference; in Adobe Firefly, build it as a style reference plus brand-kit colors; in Figma Weave, keep the block pinned for in-canvas generation and restyles. The master prompt is versioned like a design token — changes go through review, not vibes.

3. 3.

   Art-direct every batch; never prompt-and-publish. Generation runs per asset class: master prompt + subject line + format specs. Direct like you'd direct a human illustrator — good direction notes look like:

   Copy

   BATCH: empty states, 4 subjects (no results / no connection /
   first use / all done)
   Round 1: composition too dense on "no results" — more negative
   space, subject smaller. "All done" reads as celebration; brand
   register is calm — remove confetti.
   Round 2: approved 3 of 4. "First use" hands are wrong — regenerate
   hands or crop above wrists.
   

   The reject-and-redirect notes are the pipeline's quality system; keep them with the batch.

4. 4.

   Post-process into system assets. Winning generations get cleaned in your design tools — cropped to format, colors snapped to brand tokens where the model drifted, exported at the pipeline's standard sizes, filed with the naming convention pb-10's hygiene rules established. An asset isn't done until it's a library asset, not a download.

5. 5.

   Write the rights-and-disclosure policy before scale-up. One page, ratified with whoever owns legal risk:

   Copy

   Draft a one-page generative-asset policy covering:
   - TOOLS: which generators are approved, under which license tier,
     and what their terms say about commercial use and training
   - RIGHTS: what we can claim over generated output per tool;
     where human authorship must be added for protectability
   - INPUTS: what may never be used as a style reference (competitor
     art, unlicensed artists' work, client IP without contract cover)
   - DISCLOSURE: where we label assets as AI-generated (and where
     regulation or platform policy requires it)
   - PEOPLE: rules for generated human likenesses
   Blunt language. Name an owner for exceptions.
   

Make it yours

• Firefly-first for enterprise caution: if your organization needs indemnification and training-data provenance, build the pipeline on Firefly and Adobe's stack, and accept the style-range tradeoff — the policy page writes itself much shorter.
• Weave-first for in-product assets: when most volume is in-canvas (empty states, onboarding art), keeping generation inside Figma means assets land where they're used and restyles inherit file context.
• Midjourney-first for range, with a hard post-process gate: widest aesthetic range, loosest fit to brand tokens — pair it with a mandatory color-snap and cleanup step, and check its current commercial terms against your policy.

Quality check

• Ten assets side by side read as one productcoherence is checked at batch level, not per image; if the wall looks like an anthology, tighten the master prompt.
• Every published asset passed human art directiona batch with zero rejects means the bar is on the floor, not that the model is perfect.
• Colors match brand tokens exactlymodels approximate; eyedropper-verify against the palette, snap what drifted.
• Nothing traces to someone else's styleno artist names in prompts, no competitor references; near-copies of identifiable work are a rights problem AND a brand embarrassment.
• Representation was reviewed deliberatelygenerated humans inherit training-data defaults; check that the people in your assets reflect your actual audience, not the model's priors.
• The policy exists and names an ownera pipeline scaling faster than its policy is how brands end up apologizing on social media.