Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

AI in Marketing › AI Applications in Marketing

Producing AI-Generated Ad Images with Review

A step-by-step process for producing ad images with AI: brief, references, generation, selection, editing, rights and likeness checks, labels and final QA.
Tutorialintermediate9 min readUpdated

What this process covers

The process covers still images for paid placements, produced wholly or partly with a generative tool: a scene generated around a real product photograph, a background extended to fit a new aspect ratio, a supporting image for a carousel, or a fully generated illustration. It does not name tools or rely on any tool's particular features, since those change; every step works with any image generator or editor the organisation has approved.

Three things set generated images apart from photography or design, and the extra steps exist because of them. The model can render details wrongly in ways a photograph cannot, such as hands, text, logos, reflections and the product itself. The inputs, prompts and reference images, can carry rights and likeness problems into the output without anyone intending it. And the ad platforms document rules for labelling realistic generated or altered imagery in some contexts, which a photograph would not trigger.

StageOrdinary creativeGenerated image adds
BriefObjective, audience, message, mandatoriesWhat may and may not be generated; where the real product must appear
InputsShoot or stock selectionReference rights, prompt rules, a log of both
ProductionDesign and retouchingSelection from many options; compositing real and generated parts
ReviewBrand, claims, accessibilityDetail and product accuracy, likeness, platform labelling
RecordsVersion and sign-offTool, prompt, references and edits for every approved image

Everything else follows the organisation's normal process: the creative production playbook for the overall flow and the creative approval procedure for sign-off.

Before you start

  • An approved tool. One your AI policy permits for commercial creative, with terms someone has read covering commercial use of outputs and what happens to uploaded material.
  • The brand's imagery rules. The imagery section of the brand guidelines: subjects, settings, light, colour and what the brand never shows.
  • Real product assets. High-quality, approved photographs or renders of the product, packaging and logo. Generated images should never be the source of truth for what the product looks like.
  • A disclosure rule. The organisation's decision on when generated imagery is labelled, in line with the disclosure and data privacy guide.
  • Placement specifications. The sizes, aspect ratios, safe areas and file requirements for each placement the image will run in, from the platform's current help centre pages.
  • A production log. A shared sheet or asset-library fields for tool, prompt, references, edits and approver.

Producing one ad image, step by step

  1. Write the image brief

    Start from the campaign's creative brief and add a section for the image. State the message the image carries, the audience and the placement, the subject and setting, the mood, and the mandatory elements such as the product, logo and any text overlay. Then state what the tool may generate and what must be real: usually the product, packaging, logo and any text are real assets composited in, while the setting and supporting elements may be generated. List what must not appear, including people who could be mistaken for real customers, competitor products and recognisable locations the brand has no right to use.

  2. Gather references

    Collect the reference images the tool will receive: the brand's own photography, previous approved campaign images and licensed stock that matches the imagery rules. Every reference must be something the organisation owns or holds a licence that covers this use. Do not use images found online, other brands' advertising or an artist's portfolio as references. Write the prompt from the brief in plain descriptive terms, and keep the names of real people, brands, characters and artists out of it. Record the references and the prompt in the log before generating anything.

  3. Generate options

    Generate a set of options from the prompt and references, varying one thing at a time if the brief allows choice, such as the setting or the time of day. Keep the prompt and references fixed within a round, so that differences between options come from the generation and not from edits to the instructions. If none of the options is close, revise the brief or the references rather than piling adjectives onto the prompt. Save the options with the round number and the prompt that produced them.

  4. Select against the brief

    Lay the options out together and compare each with the brief, not with each other. Does it carry the message, fit the imagery rules and leave room for the product, logo and text in the safe area of every placement? Reject anything that only looks impressive. Choose a small shortlist and record why each was chosen. If the image is part of an A/B Test, select options that differ only in the element the test names, so the result can be read.

  5. Edit and composite

    Bring the shortlisted image into an editor. Composite the real product photograph, packaging and logo into the scene rather than relying on generated versions of them. Match light direction, colour temperature, shadows, reflections and scale so that the real and generated parts sit together naturally. Add text overlays as live text in the design tool, never as text rendered by the generator. Extend or crop for each aspect ratio the placements need, and record every edit in the log.

  6. Check details and product accuracy

    Review the edited image at full size and at the size it will be seen on a phone. Check hands, faces, fingers, teeth and eyes; any text or signage in the scene; logos and labels; reflections, shadows and edges where composited parts meet; repeated or impossible objects; and the physics of the scene. Then compare the product against the real thing: shape, colour, label, size relative to the surroundings and anything shown in use. An image that makes the product look larger, different or capable of something it is not is a misleading claim, not a styling choice.

  7. Check rights and likeness

    Confirm that every reference was owned or licensed for this use, that the prompt named no real person, brand, character or artist, and that the tool's terms permit commercial use of the output. Look hard at any generated person: does the face resemble a real, recognisable individual, a celebrity or a member of staff? If so, replace it. Check that no generated element reproduces another brand's mark, packaging or a recognisable work. Ownership of generated material is an open legal question in many places, so record what is known and escalate anything uncertain to whoever advises the organisation; the copyright and training data questions set out what is unsettled, without legal conclusions.

  8. Label and export for each placement

    Check the platform's current help centre pages for the campaign's category at upload time. Both major platforms document requirements to disclose or label digitally created or altered media in at least some categories, political and social-issue advertising most clearly, and some apply their own labels to content they detect as generated. Apply the platform's label where required and the organisation's own disclosure rule otherwise. Export each placement size to the specification, check every crop on a real device, and save the approved files with version labels. If the platform offers automatic enhancements, preview what they change and decide whether to keep them.

The review gate

Before the image goes to the approval procedure, one reviewer who did not produce it signs off against this gate. Failures go back to the step that caused them, not to a quick fix in the ad platform, because a fix made outside the log is invisible later. If an image has already been rejected or labelled by a platform after upload, the diagnosis for rejected AI-generated ad creative works back from the platform's reason to the gate check that was missed.

CheckPass whenTypical failure
Brief fitThe image carries the brief's message for the stated audience and placementA striking image that says something else
BrandSubjects, setting, light and colour follow the imagery rulesThe generic polish of an unbranded generated image
Product accuracyThe product matches the real one in shape, colour, label and scaleA generated product with the wrong cap or proportions
DetailsHands, faces, text, logos, edges and reflections are correct at full sizeExtra fingers, warped lettering, a shadow from the wrong side
RightsReferences owned or licensed; no names in the prompt; tool terms permit the useA reference copied from another brand's campaign
LikenessNo generated person resembles a real individualA face that looks like a known presenter
LabellingPlatform rules checked at upload; disclosure rule appliedA realistic altered image in a category that requires a label
Placement fitEvery size checked on a device with safe areas respectedLogo cropped out of the vertical version
LogTool, prompt, references, edits and approver recordedAn approved image nobody can trace

After launch, judge the image on the metric the campaign is bought for, not on CTR alone; the CTR calculator helps compare click-through across versions, but the version that earns the most clicks is not always the one that sells. Generated images can tire as quickly as any other creative, so watch for Creative Fatigue and plan refreshes through the same steps rather than by regenerating at random.

Illustrative example — not a real result

Producing a seasonal ad image for a candle maker

Kestwick Candles, an invented candle maker, needs an autumn image for social feeds in 3 aspect ratios. The brief says the candle, its label and the logo must be real assets; the setting, a windowsill at dusk, may be generated. The designer uses 6 of the brand's own photographs as references and generates 24 options in 2 rounds. 5 are shortlisted; 3 are rejected at the detail check because of warped window frames and a second, impossible shadow behind the candle. The chosen image has the real product photograph composited in, with its glow redrawn to match the scene's light. The likeness check finds nothing, since no people appear. The image is exported in all 3 ratios, and the vertical crop is redone because it cut the label in half. Review and fixes take about 2 hours on top of generation, and the log records the tool, the prompt, the 6 references and the approver's name.

Common mistakes

  • Generating the product. The real product should be composited in; a generated one drifts from what customers receive.
  • Using references the team does not own. Rights problems usually enter through the inputs.
  • Names in the prompt. People, brands, characters and artists in a prompt create likeness and rights exposure.
  • Reviewing at thumbnail size. Warped hands, text and edges only show at full size.
  • Letting the generator render text. Overlays belong in the design tool as live text.
  • Checking labels from memory. Platform rules change; check the help centre at upload time.
  • Fixing problems in the ad platform. Fixes outside the log leave no record of what ran.
  • No log. Without the tool, prompt, references and approver recorded, a later rights or disclosure question has no answer.

Frequently asked questions

Do I have to label AI-generated images in ads?

It depends on the platform, the ad category and how realistic the image is, and the rules change. Both major ad platforms document labelling or disclosure requirements for digitally created or altered media in at least some categories, most clearly political and social-issue advertising, and some apply their own labels. Check the platform's current help centre page at upload time, follow your organisation's disclosure rule, and never use a generated person to imply a real customer or employee.

Can I use AI to put our product into a generated scene?

Yes, and compositing a real product photograph into a generated or extended background is usually safer than generating the product itself, because the product stays accurate. Check the edges, shadows, reflections and scale where the two meet, and confirm that nothing about the product's shape, colour, label or size has changed. An image that shows the product differently from what customers receive misleads them, however it was made.

Who owns an AI-generated ad image?

That is an open question that depends on the tool's terms and on how copyright law in each jurisdiction treats generated material, which is unsettled. What the team can control is the input side: use only references it owns or has licensed, keep names of people, brands and artists out of prompts, read the tool's commercial terms, and keep a log of how each image was made. Take advice on anything with commercial weight.

Terms used on this page

Tools to try

Where this page is used

Also related

Sources

  1. Meta Business Help Centre — Meta
  2. Google Ads Help — Google Ads Help

Related