AI Image Generation: Briefs, References and Selection

#01What job should the image do?

Start with the communication problem. A campaign image might create a recognisable visual theme, explain an abstract process or provide background space for a headline. It should not be generated merely because a channel has an empty image slot. Define the message a viewer should understand and the role the image plays alongside real text.

Classify the image as conceptual or evidentiary. A stylised illustration of document routing can be conceptual. A screenshot, customer photograph or event image usually implies something factual. Generating a realistic substitute can mislead viewers even if the caption does not explicitly make a false claim.

Decide where genuine assets are necessary. Product interfaces should generally come from the real approved interface or a clearly labelled conceptual mockup, not a generated screen presented as working software. Customer stories need properly authorised evidence, not synthetic people implying a real endorsement.

Set a non-generation option. Licensed stock, commissioned illustration, a diagram or an existing approved asset may communicate more clearly. The brief should permit that choice if generated candidates cannot satisfy the intended use responsibly.

#02How do you write a production-ready visual brief?

Describe the subject, setting, composition, visual treatment and practical constraints. Explain the image's purpose and where real headline text will sit. Avoid a long pile of style adjectives that gives no guidance about hierarchy, meaning or the viewer's attention.

Specify what must be absent. For a conceptual software campaign, exclusions might include customer logos, invented dashboards, readable financial records or people appearing to endorse the product. These are editorial constraints, not a claim that a negative instruction will reliably prevent every defect.

Include delivery context: intended placements, crops, text-safe areas, background compatibility and accessibility considerations. Do not assume a candidate that works in a wide composition will survive a narrow crop. Test the actual placement rather than relying on the generation preview.

Use plain language and iterate one important variable at a time. If subject, palette, camera angle and composition all change together, the team cannot learn which change improved the result. Save useful briefs with the selected asset rather than leaving them in an untraceable chat.

#03What should you check about models and permissions?

Adobe Firefly is both an application environment and a family of capabilities. The current FAQ describes Adobe's own model practices and other product contexts. Do not assume that every partner model available through an Adobe interface has identical training, terms or protections.

Record the actual model used and relevant account conditions. A workflow approved for one model is not automatically approved for a newly selected partner model. The model picker is an operational dependency, not a decorative setting.

The checked Firefly FAQ permits commercial use of non-beta outputs and generally beta outputs unless a product says otherwise. That is not a guarantee of copyright protection, non-infringement or indemnification for every account and use. Review the applicable terms and escalate high-stakes rights questions appropriately.

References also need permission. A publicly visible photograph is not automatically available for upload, transformation or style reference. Check ownership, licence, subject permissions and any client restrictions. Where rights are unclear, use an approved alternative rather than treating uncertainty as a creative inconvenience.

#04How should references be used responsibly?

Explain what a reference contributes: colour balance, composition, lighting or a high-level visual principle. Do not use another brand's distinctive assets as an instruction to create a confusing imitation. A useful brief can describe the desired effect without asking for a counterfeit identity.

Keep source images separate from generated outputs in the asset register. Record permission and intended processing. A customer photograph licensed for a named case study may not be approved for synthetic variants, unrelated markets or a new campaign.

Avoid presenting a synthetic person as a real employee, buyer or customer. If people are necessary, consider whether an illustration, genuine authorised photography or an explicit synthetic treatment better fits the message. Disclosure and context should make the nature of the asset understandable where relevant.

Do not treat reference fidelity as a guaranteed product capability. Features and controls vary by model and plan. Inspect the actual result, especially identity, logos, product geometry and text, rather than assuming that a supplied reference will be preserved exactly.

#05How do you select candidates without rewarding spectacle?

Write the rubric before reviewing the contact sheet. Score communication fit, composition, truthful implication, brand suitability, editability and delivery readiness. Keep rights uncertainty and misleading factual representation as rejection gates rather than weaknesses that beauty can outweigh.

Review at several scales. A thumbnail reveals whether the central idea reads quickly. A larger view reveals malformed objects, unreadable text and distracting artefacts. The final placement reveals crop and contrast problems. A beautiful full-frame preview can fail every real channel treatment.

Record why candidates are rejected. "Wrong mood" is less useful than "implies a real payment dashboard we do not offer" or "leaves no usable headline area". Clear rejection reasons improve the next brief and help distinguish tool limitations from unclear direction.

Do not report a hypothetical selection as a measured tool benchmark. If a team later runs a real evaluation, retain all attempts and review costs. This lesson's example is a production plan, not evidence that one model is more effective than another.

#06What belongs in the final asset handoff?

Deliver the selected file, source brief, actual model, relevant references, rights record, significant edits and approved uses. Include the final export and any editable working files needed by the next operator. A download alone is not a complete handoff.

Adobe's Content Credentials can communicate provenance where supported, but provenance is not proof of factual truth. A credentialed image can still imply an event that never happened. Keep the communication review independent of the presence or absence of a provenance indicator.

Separate commercial-use permission from copyrightability. The U.S. Copyright Office's Part 2 report explains that protection depends on human authorship and the particular contribution; prompts alone do not generally provide sufficient control under the analysis described there. That is U.S. guidance, not a universal legal conclusion for every jurisdiction.

Use the existing ChatGPT image-generation lesson for vendor-specific practice. This lesson's distinct purpose is the cross-tool production brief, selection discipline and rights-aware handoff.

#07Worked example: Ledgerline

Ledgerline needs an illustrative campaign visual about reducing the confusion around invoice approvals. The team chooses an abstract editorial treatment rather than a fabricated screenshot or synthetic customer photograph. The following brief and review record are illustrative.

The team prepares a contact-sheet review with three possible rejection categories: misleading representation, weak communication and delivery problems. It does not invent a customer outcome or claim the generated image proves a product benefit.

Walkthrough

1. Product marketing approves the communication goal and states that the asset must not resemble evidence of a real customer workflow.
2. Design selects permitted references and explains the visual principles they contribute.
3. The operator checks the actual Firefly model and relevant account terms before creating candidates.
4. Reviewers inspect the contact sheet against the rubric, recording specific defects rather than choosing only by preference.
5. The selected candidate is placed with real text in the intended crops and checked for misleading implications, contrast and distracting details.
6. The final handoff includes the brief, rights record, significant edits, approved uses and source files needed for future adaptation.

If no candidate meets the truth boundary, the fallback is a commissioned or manually designed illustration. The production goal is useful communication, not forcing generation into every campaign.

#08Common mistakes

• Generating a fake product screen. Use real evidence or clearly labelled conceptual treatment.
• Treating all models in one interface as identical. Check the actual model and terms.
• Using references without permission. Visibility does not establish upload or transformation rights.
• Selecting by visual impact alone. Review communication, truthful implication and delivery fit.
• Forgetting text placement and crops. Test the image in its real campaign context.
• Calling provenance a truth certificate. It documents aspects of origin, not the accuracy of the scene.

#09Try it yourself

Write a visual brief for one real campaign need. Specify the communication role, subject, composition, approved references, exclusions, truth boundary and delivery contexts. Include a non-AI fallback and identify which rights questions must be resolved before production.

Create or inspect a small permitted contact sheet. Apply a rubric written in advance and record at least one substantive rejection reason. Do not invent comparative performance figures if you have not run a controlled test.

Deliver the selected concept with its brief, rights record and placement review. Ask a colleague to explain what the image appears to claim. If that inference exceeds the approved message, revise the visual or choose another approach.

#10Quick quiz

1. Does commercial-use permission guarantee copyright protection?

Answer: No. Permission, copyrightability and other rights questions are separate.

2. What should happen when a visually strong image implies a false product capability?

Answer: Reject or revise it; aesthetic quality cannot override the truth boundary.

3. Why record the actual model rather than only the application name?

Answer: Different models can have different capabilities, terms and operating conditions.

#11What should you check before moving on?

Confirm the selected asset communicates the intended idea without false evidence, has permitted inputs and survives actual placements. Continue with image-editing and design QA before treating a generated candidate as a finished campaign asset.

Previous: AI Copy and Localisation: Terminology, Meaning and QA · Next: AI Image Editing and Design QA Without Product Distortion

Or return to the AI Tools for Marketers track.

Related: Brand Controls for AI Work: Context, Templates and Review · AI Image Editing and Design QA Without Product Distortion · AI Content Rights, Disclosure and Human Editorial Review

#12Common questions

What should an AI image brief include? Communication role, subject, composition, permitted references, truth boundary and delivery context.

Does Firefly commercial use guarantee copyright? No. Commercial-use eligibility and copyrightability are separate questions.

Can a generated screen be used as product evidence? Not if it invents the interface or behaviour. Use real approved evidence or clearly labelled conceptual treatment.