A Worked Example of AI Image Generation: From Rough Idea to Usable Creative Material

The failed first pass: a beautiful image that cannot ship

Consider a common brief: “Create a premium key visual for an original fantasy IP collaboration page—young hero, spirit wolf, neon rain, elegant city, cinematic.” A generator can produce something impressive in minutes. The first pass may still be useless.

The hero's coat changes shape between variants. The wolf looks like an ordinary husky rather than a recognizable companion. The city contains pseudo-signage that resembles brands. There is no safe space for bilingual copy. The lighting is dramatic but hides the product area where a sponsor logo will eventually sit. Nobody recorded which reference image influenced which decision.

The mistake is not a weak prompt. The mistake is starting production before defining the deliverable.

Rewrite the job as an art-direction brief: 16:9 hero image for web; hero occupies lower-right third; spirit wolf must preserve white coat, ear silhouette and scale relationship; upper-left stays quiet for bilingual headline; no readable third-party marks; wet-night palette but faces remain legible; final asset needs an editable composition plan and a record of references and human edits.

Now generation has a target beyond “looks good.”

Separate reference roles before touching the model

The team has six references: an approved character sheet, a wolf turnaround, a rainy street photo, a luxury editorial layout, a material board and an old campaign image. Uploading all six without labels invites drift.

Classify them instead.

Reference	Job	Must preserve	Must not borrow blindly
character sheet	identity	face cues, coat construction	original background
wolf turnaround	identity/scale	ears, white coat, body ratio	pose
rainy street	atmosphere	wet reflections, haze	signage, people
editorial layout	composition	negative-space logic	brand-specific styling
material board	surface	metal/cloth contrast	exact object shapes
old campaign	constraint	sponsor-safe area	obsolete product details

Some current tools expose separate controls for style and structure references; Adobe's Firefly documentation, for example, describes distinct style-reference and structure-reference concepts. The exact controls vary by product and version, so the durable practice is conceptual: label why a reference exists before you use it.

Rights are a separate gate. An accessible image is not automatically an image you are entitled to upload, transform or publish from. Check source rights, client permissions and current tool terms.

Generate for composition before texture

The second pass should answer one question: does the image read correctly at thumbnail size?

Ignore embroidery, rain droplets and micro-texture. Generate or block several compositions with the hero and wolf as simple masses. Reject anything that steals the headline zone, collapses the wolf into the background or puts the characters in a crop that will fail on mobile.

Choose one layout and record why: “Version C keeps the hero's face inside the safe crop, gives the wolf a clean silhouette against a darker wall, and leaves 34 percent of the upper-left visually quiet.”

The percentage does not need to become a universal rule. The value is traceability. Another designer can understand what the selected layout is protecting.

Then move to identity. Compare the approved character anchors, not vague resemblance. Is the hair direction right? Are the coat closures in the right place? Is the wolf's ear geometry recognizable? If identity is still drifting, do not spend time polishing puddle reflections.

A good workflow spends detail only after the expensive-to-change decisions stabilize.

Run a controlled iteration log

Once composition and identity are credible, iteration becomes a sequence of controlled questions.

Pass A asks whether spatial logic is believable: hands, weapon grips, paw placement, stairs, reflections, furniture and perspective. Pass B asks whether material behavior supports the design: wet cloth should not look like chrome; brushed metal should not melt into skin. Pass C tests delivery conditions: desktop crop, mobile crop, text overlay, dark-mode surroundings, compression and thumbnail recognition.

Record each pass in one line: problem → change → result → keep/reject.

This prevents prompt inflation. Without a log, teams often add adjectives—“more cinematic, more premium, more detailed”—until nobody knows which change improved the image. A controlled pass changes one family of variables and keeps the rest as stable as practical.

If a local defect needs substantial manual repair, compare the repair time with other methods. A conventional paint-over, 3D block-in, photography or compositing step may be faster and more controllable than another thirty generations.

Human authorship, provenance and the review package

For commercial work, save more than the final JPEG.

Keep the approved brief, important reference permissions, selected intermediate versions, edit files and a short note describing human contributions such as layout changes, compositing, masking, paint-over, typography and final selection. In the United States, the Copyright Office's 2025 AI copyrightability report emphasizes that copyright depends on sufficient human-authored expressive elements and that prompts alone do not automatically provide that authorship. The legal result always depends on the actual work, so documentation is useful without being a guarantee.

Where a workflow supports Content Credentials or other C2PA-based provenance, understand the boundary: provenance can record assertions about origin and editing history. It is not a machine that certifies an image is true, safe, licensed or aesthetically good.

For this example, the review package contains the brief, reference-role table, iteration log, the layered final file, export settings and provenance information available from the chosen tools. That package is much easier to hand to a partner than “here is the image we liked best.”

The final preflight: usable means it survives the real page

Place the asset in a real page mockup before calling it finished. Add the actual bilingual headline, navigation, CTA and partner mark. Test the responsive crops. Check the image against the site's compression settings. View it on a bright screen and a dim one.

Then do a claims pass. If the visual depicts a product, place, uniform, historical object or branded feature, ask whether viewers could treat the image as a factual representation. A fantasy illustration can be highly stylized; a product image used next to a buy button carries a different risk of misleading detail.

Finally, do a red-team glance by someone who did not participate in generation. Ask them to point out accidental marks, anatomy failures, cultural symbols, unsafe visual implications and places where the image contradicts the page copy.

The finished asset is therefore not “the generation.” It is the result of specification, reference discipline, controlled iteration, human editing, rights/provenance review and delivery testing. That is the difference between a rough AI image idea and a usable creative asset.

What this worked example teaches

The worked example can be reduced to a reusable sequence: define the delivery surface, label references by function, solve composition, stabilize identity, iterate one question at a time, document human editing, review rights and provenance, and test the asset in context.

None of those steps requires a particular model. That is intentional. Image-generation systems change quickly; controls that are available in one product or release may not exist in another. Treat current documentation as the authority for tool behavior.

The durable skill is art direction under uncertainty. A useful team can explain what must remain stable, what is allowed to vary, how a reviewer will recognize failure, and when generation is no longer the cheapest tool for the job.

That is also the best defense against the “prompt magician” trap. The valuable output is not a mysterious sentence that happened to work once. It is a process another competent person can inspect, repeat and improve.

A decision table for the final selection

Before choosing the winner, score the surviving versions against the brief rather than against each other.

Criterion	Weight in this job	Version A	Version B	Version C
hero and wolf identity	high	weak	strong	strong
headline-safe negative space	high	strong	weak	strong
mobile crop	high	weak	medium	strong
product/sponsor area	medium	strong	medium	strong
spatial plausibility	medium	medium	strong	strong
repair effort	medium	high	low	medium

The numbers do not have to become a pseudo-scientific score. Their purpose is to stop one spectacular feature from hiding a fatal delivery problem. A version with extraordinary rain lighting but no mobile crop is not “almost done”; for this brief it is a different concept.

Version C wins because it clears the important constraints with manageable repair work. The rejected images are not wasted. Tag them by the thing they did well—lighting, atmosphere, pose—and keep them as internal references if the team has the right to retain them.

This selection step also creates a clean approval conversation. Instead of “I like B more,” the art director can say, “B has the best face, but C is the only one that preserves identity, text space and mobile crop simultaneously; we will borrow B's face treatment during the local edit.” That is a decision a production team can execute.

How the case handles a stakeholder change

Halfway through production, imagine the partner asks for a square social crop and requests the wolf to become more prominent. A weak workflow starts over because the original image was optimized as a single flattened composition.

The controlled workflow returns to the brief and asks whether the request changes the hierarchy or only the delivery. If the wolf must now share primary focus, that is a hierarchy change. Re-open composition before texture. Make a square blockout, protect the character anchors, and reuse only assets that survive the new relationship.

If the request is merely “we also need a square derivative,” keep the hierarchy and build a deliberate alternate crop or extension instead of stretching the original.

This distinction—new concept versus new rendition—saves substantial rework. It also belongs in the handoff notes, because licensing, approval and provenance records may need to follow the derivative asset rather than disappearing when the file is reformatted.

A final naming convention that keeps the work usable

Name approved files by campaign, asset role, aspect ratio and version rather than by prompt mood. A filename such as collab-hero-desktop-16x9-v07-approved tells the next person what they are opening. Preserve the editable source beside the delivery export. Good file hygiene is not glamorous, but it prevents a later designer from publishing the wrong “final-final-2” image after all the careful review above.

Keep a rejected-version library with reasons

Do not keep every generated image forever, but preserve a deliberate subset of rejected versions with rejection reasons. One may have the correct wolf identity but unusable text space; another may have excellent atmosphere but incorrect coat construction.

Those labeled failures are valuable training material for the team. They make future briefs sharper and help reviewers distinguish “aesthetic preference” from a recurring technical problem.

Avoid turning the folder into an ungoverned archive of third-party references or confidential client material. Retention should follow the permissions and information-handling rules that apply to the project.

A small, annotated failure library often teaches more than a gallery containing only polished winners, because it preserves the decisions that the final image no longer reveals.