AI Image Generation and Design · Consistency across a campaign · lesson 12 of 18 · 7 min

A production pipeline from brief to delivery

Scaling without chaos

When AI imagery moves from occasional experiments to regular production — weekly content, ad variations, campaign launches — you need a pipeline: clear stages, owners, checkpoints and file conventions. This protects quality and makes the process repeatable.

The pipeline

| Stage | Activities | Output | Owner | |---|---|---|---| | 1. Brief | Objective, audience, formats, message, constraints, where AI is/isn't allowed | Creative brief | Strategist/client lead | | 2. Concept | Moodboards, AI explorations, rough layouts | 2–3 concepts | Designer | | 3. Approval of direction | Client/stakeholder picks concept | Approved direction | Client | | 4. Generation | Using style system, references, prompt log | Candidate images | Designer | | 5. Selection | Shortlist against brief | Selected images | Designer + art director | | 6. Editing & compositing | Inpaint, outpaint, composite, retouch, upscale | Finished images | Designer/retoucher | | 7. Layout | Typography, logos, templates, formats | Final assets | Designer | | 8. QA & compliance | Artifacts, brand, rights, bias, claims, disclosure | QA sign-off | Reviewer | | 9. Delivery | Exports, naming, metadata, documentation | Delivered package | Designer/PM | | 10. Learn | Performance review, update system | Updated style system | Team |

Small teams may combine roles, but every stage still happens.

The brief: AI-specific questions

Add these to your standard creative brief:

AI-specific brief questions
- Is AI imagery acceptable for this client/campaign? Any client policy?
- Which elements must be real (products, people, places, results)?
- Are there sensitive topics, cultural contexts or regulated claims?
- Required disclosures (platforms, markets, client policy)?
- Tool restrictions (data privacy, commercial terms, indemnity needs)?
- Ownership expectations for final assets?

Ad variations at scale

AI makes it easy to produce many variations for testing. Keep it disciplined:

• Vary one element at a time (background scene, model pose, color of backdrop).
• Keep brand elements fixed through templates.
• Name files so test results can be traced back: campaign_placement_variable_version (e.g. launch_story_bg-desert_v2).
• Retire underperforming concepts quickly; iterate winners.
• Check that each variation is still compliant — variations can introduce new errors.

Delivery package

Delivery package
/Final_exports     platform-ready files (named, sRGB, correct sizes)
/Masters           layered working files
/Source_images     AI outputs used, real photos used (with licenses)
/Documentation     prompt logs, references used, AI usage summary,
                   disclosure recommendation, rights notes

An AI usage summary — a short note of which assets used AI and how — helps clients with their own disclosure and governance obligations.

Measuring and learning

After publication, review:

• Performance metrics suited to the objective (engagement, click-through, conversions, watch time).
• Qualitative feedback (comments noting "AI look" or confusion).
• Production metrics: time and cost per final asset, number of revisions.
• Update the style system, prompt snippets and gallery with what worked.

Worked example: monthly content for a UAE restaurant group

• Brief each month includes which dishes must be real photography (all menu items) and where AI can be used (seasonal backgrounds, illustrated Ramadan and Eid graphics, abstract textures).
• Generation uses the group's style system; seasonal snippets are added for Ramadan (lanterns, crescent motifs, warm evening light) reviewed by team members for cultural appropriateness.
• QA checks that no generated food is presented as menu items and that Arabic text is set in the design tool.
• Delivery includes an AI usage summary per asset.

Common mistakes

• Skipping direction approval and generating hundreds of finals in the wrong style.
• No QA stage for compliance and disclosure.
• Losing track of which variation is which.
• Not documenting AI use for clients.

Tools for managing the pipeline

You do not need specialist software to run this pipeline. A shared task board with a column for each stage, a naming convention, and a folder template are enough for most small teams. What matters is that each asset has an owner, a status and a place where its prompt log, references and approvals live. As volume grows, consider digital asset management tools that store metadata such as AI usage and license information alongside each file, making audits and client questions much easier to answer.

Hands-on: automate variants with an image API (optional)

For teams producing many variants, an API call can generate candidate backgrounds from your library. This example uses the official OpenAI Python SDK's Images API; the model name is read from an environment variable because models change — check the current model list in the vendor's documentation. The same pattern (prompt in, file out, log entry written) works with other vendors' APIs.

import base64
import json
import os
from datetime import date

from openai import OpenAI  # pip install openai; needs OPENAI_API_KEY in the environment

client = OpenAI()
MODEL = os.environ.get("IMAGE_MODEL", "gpt-image-1")  # check current model names

jobs = [
    {"asset_id": "eid-bg-01", "prompt": "Empty warm-lit courtyard with lanterns at dusk, "
     "calm top third for a headline, no people, no text, no logos", "size": "1024x1536"},
]

for job in jobs:
    try:
        result = client.images.generate(model=MODEL, prompt=job["prompt"], size=job["size"])
        image_bytes = base64.b64decode(result.data[0].b64_json)
        path = f"out/{job['asset_id']}.png"
        os.makedirs("out", exist_ok=True)
        with open(path, "wb") as fh:
            fh.write(image_bytes)
        log = {**job, "model": MODEL, "date": date.today().isoformat(), "file": path,
               "status": "needs_review"}  # a human reviews every output before use
        with open("out/prompt_log.jsonl", "a", encoding="utf-8") as fh:
            fh.write(json.dumps(log) + "\n")
    except Exception as err:  # log and continue; never retry blindly in a loop
        print(f"{job['asset_id']}: generation failed: {err}")

Outputs from this and other major vendors typically carry C2PA provenance metadata; keep it through editing and export where your tools support it.

Delivery package template

/ClientName_Campaign_2026-10/
  /01_final_exports/      (per platform, named asset_platform_size_v#.ext)
  /02_masters/            (layered files: .psd/.fig links, .ai)
  /03_sources/            (product photos, licensed images + license PDFs)
  /04_ai_record/          prompt_log.jsonl, references, AI_USAGE_SUMMARY.md
  /05_qa/                 signed QA sheets, disclosure decisions per asset

The AI usage summary is one page: tools and models used, which assets contain AI-generated or AI-edited elements, which elements are real, disclosure recommendations per platform, and known limitations (for example, "purely AI-generated backgrounds may have limited copyright protection").

Summary

Run AI imagery through a staged pipeline with owners and checkpoints, add AI-specific questions to briefs, keep variation testing disciplined, deliver organized packages with an AI usage summary, and learn from results.

Video lecture: A production pipeline from brief to delivery

Lecture coming soon · 10 chapters · about 8 minutes. Read the full transcript below.

1. A production pipeline
2. Why a pipeline
3. The ten stages
4. AI questions for every brief
5. Worked example 1: a bakery's weekly posts
6. Worked example 2: a five-branch restaurant group
7. Watch me do it: the delivery package
8. Optional automation
9. Common mistakes
10. Recap and try this now

Lecture transcript

A production pipeline

Here's the moment AI imagery gets serious. A client stops asking for one hero image and starts asking for forty assets a month, across four platforms, in three languages, with ad variations to test. If your process is a designer and a chat window, it will break. Things get lost, versions get mixed up, someone publishes an unreviewed image, and nobody knows which assets used AI. In this lecture, you'll learn a ten-stage production pipeline, the AI-specific questions to add to every brief, how to produce ad variations at scale, and what goes into a professional delivery package. By the end, you'll be able to run AI image production like a studio, not a hobby.

Why a pipeline

Why a pipeline? Because at scale, the risk isn't one bad image. It's a system that can't tell good from bad, reviewed from unreviewed, or real from generated. A pipeline gives you stages, owners, checkpoints and file conventions. Think of an airport. Hundreds of planes land safely every day, not because pilots are perfect, but because there are checklists, handoffs and clear responsibilities at every step. Your pipeline does the same for images. It also makes your costs predictable, so you can price the work properly and protect your margins.

The ten stages

Here are the ten stages. One, brief: objective, audience, formats and where AI is or isn't allowed. Two, concept: moodboards, AI explorations and rough layouts. Three, approval of direction by the client. Four, generation, using the style system and prompt log. Five, selection against the brief. Six, editing and compositing: inpaint, expand, composite and upscale. Seven, layout with typography, logos and templates. Eight, QA and compliance: artifacts, brand, rights, bias, claims and disclosure. Nine, delivery: exports, naming, metadata and documentation. And ten, learn: review performance and update the system. Small teams combine roles. But every stage still happens.

AI questions for every brief

Stage one deserves extra attention, because most problems start with a brief that never mentioned AI. Add these questions to your standard brief. Is AI imagery acceptable for this client? Is there a client AI policy? Which elements must be real, like products, people, places or results? Are there sensitive topics, cultural contexts or regulated claims, like health or finance? What disclosures are required on each platform and in each market? Are there tool restrictions for confidentiality? And who approves AI-assisted assets? Here's the key idea. Five minutes of questions at the start prevent five days of rework at the end.

Worked example 1: a bakery's weekly posts

Worked example one, simple. A freelancer makes weekly posts for a local bakery. She keeps the pipeline light. The brief is a shared note with the AI questions answered once: AI allowed for backgrounds and seasonal scenes, never for the actual cakes. Every Monday she generates backgrounds from her snippet library, composites real cake photos taken on her phone, lays them out in a template, and runs a five-minute QA with the bakery owner over a message. Delivery is a folder with exports and a one-line AI note. It's ten stages, compressed into one person and a Monday morning. The structure is still there. That's what keeps it safe.

Worked example 2: a five-branch restaurant group

Worked example two, a business scenario with illustrative numbers. A marketing team in Dubai runs content for a restaurant group with five branches. They need about sixty assets a month, in Arabic and English, across Instagram, TikTok and Google Business Profile, plus ad variations. They assign owners. A strategist owns briefs. Two designers own generation, editing and layout. A reviewer owns QA and disclosure. The account lead owns delivery. For ads, they vary one element at a time, like background scene or headline, while keeping the product photo and offer identical, so test results are clean. And every asset gets a QA sheet and a disclosure decision before it goes live. After three months, they cut the variants that never win and put that time into better photography.

Watch me do it: the delivery package

Watch me do it. I'll set up the delivery package for a campaign. I create one top folder named with the client, campaign and month. Inside, five folders. Final exports, named asset, platform, size and version. Masters, with layered files or links to Figma. Sources, with product photos, licensed images and their license documents. AI record, with the prompt log, the references and a one-page AI usage summary. And QA, with signed QA sheets and disclosure decisions per asset. Then I write the AI usage summary: tools and models used, which assets have AI elements, which elements are real, disclosure recommendations, and known limitations, like limited copyright protection for purely generated backgrounds.

Optional automation

For teams with some technical skills, part of the pipeline can be automated. The lesson includes a short Python example that calls an image API to generate candidate backgrounds from your prompt library, saves each file, and writes a log entry marked needs review. Notice two things in that script. The API key comes from an environment variable, never pasted into the code. And every output is marked for human review, because automation should speed up generation, not skip quality control. Outputs from major vendors typically carry C2PA provenance metadata, so keep it through editing and export. The same pattern works with other vendors, and with automation tools like n8n, Make or Zapier.

Common mistakes

Common mistakes. Skipping the AI questions in the brief. Letting generation start before the client approves a direction. Testing ad variants that change three things at once, so you can't learn anything. Publishing without a QA sheet because the deadline is tight. Delivering only final JPEGs, so the client can't edit or trace anything. Forgetting the license documents for stock images. And never closing the loop. Stage ten is where you look at performance, compare AI-assisted assets with other assets, and update the system. Without it, the pipeline never gets better.

Recap and try this now

Let's recap. At scale, you need a pipeline: ten stages with owners, from brief to learning. Add AI questions to every brief, get direction approved before generating, vary one element at a time in ad tests, and run QA and disclosure checks on every asset. Deliver a package with exports, masters, sources, an AI record and QA sheets, plus a one-page AI usage summary. Try this now. Map your own workflow onto the ten stages and find the one stage you're currently skipping. Then create the delivery package folder template from the lesson, and write your first AI usage summary for a recent project.

Key takeaways

• A ten-stage pipeline with named owners makes AI image production safe and repeatable at scale.
• Add AI-specific questions to every brief: acceptability, what must be real, sensitive claims, disclosure, tool limits and approvers.
• Vary one element at a time in ad tests while keeping product and offer identical.
• Deliver exports, masters, sources, an AI record and QA sheets, with a one-page AI usage summary.
• Automation (APIs, workflow tools) should speed generation, never skip human review; keep keys in environment variables.

Try it

Map your own AI image workflow onto the ten-stage pipeline, identify one missing stage, and create a delivery-package folder template with an AI usage summary document.

• Previous: Style systems and campaign prompt libraries
• Next: Copyright and ownership of AI images
• All lessons of AI Image Generation and Design