Three conclusions before the workflow

Three conclusions make the workflow easier to remember.

First: research before prompting. Collect delivery constraints, approved identity references, rights information and factual requirements before generating.

Second: draft in layers. Solve composition before identity polish, identity before local detail, and local detail before delivery finishing. Do not spend expensive effort on decisions that a later crop will invalidate.

Third: review evidence, not vibes. Every approval should point to a brief, reference, factual source, visual test or documented human decision. “It feels better” can be a useful instinct, but it is a weak production record.

The rest of the workflow explains how to operationalize those three ideas without tying the team to one specific model.

Research phase: define the asset's truth conditions

Before image research, decide what kind of truth the asset is expected to communicate.

A mood illustration can invent architecture freely. A historical reconstruction has a different burden. A product concept must be labeled if details are not final. An e-commerce product image placed beside specifications should not casually hallucinate ports, seams, controls or included accessories.

Create a “truth conditions” note:

  • fictional / illustrative;
  • concept, not final product;
  • factual product representation;
  • editorial reconstruction;
  • documentary claim;
  • internal exploration only.

Then gather sources at the level appropriate to that claim. For factual products, use approved specs and product photography. For cultural or historical material, prefer reliable primary, museum or academic sources where available. For fictional IP, use controlled canon references.

This single classification prevents many downstream arguments because the team knows whether plausibility is enough or accuracy is required.

Draft phase one: make composition cheap

The cheapest draft is not the prettiest one. Use rough generations, sketches or blocks to answer framing questions.

Test aspect ratios and responsive crops early. Put placeholder headline boxes and CTA regions directly on the mockup. If the asset will support both English and Chinese, design for the longer or denser text reality rather than an imaginary empty page.

Keep a contact sheet with rejected layouts and one-sentence reasons: “headline collision,” “mobile crop loses face,” “secondary figure competes with product,” “background horizon cuts through logo.”

This is evidence for later. When a stakeholder asks why the hero is not centered, the team can explain the system constraint instead of reopening aesthetic debate from zero.

Draft phase two: stabilize identity and structure

After layout selection, define identity anchors. For a character, that may be face proportions, hair silhouette, costume construction, key prop and scale. For a product, it may be dimensions, control placement, material transitions and logo position.

Do not ask for “more consistent.” Mark what consistency means.

Where tools support reference controls, use them deliberately and consult current documentation. A structure reference and a style reference solve different problems in products that distinguish them; they should not be assumed to behave identically across systems.

When the output drifts, decide whether to regenerate, inpaint/edit locally, composite from approved elements, or leave the AI path entirely. The workflow is successful when it reaches a controlled result—not when it proves that every pixel was generated by the same technique.

Review phase: five evidence gates

Run five gates in order.

  1. Visual hierarchy: does the intended subject win at a glance?
  2. Identity/structure: do controlled anchors match approved references?
  3. Factual integrity: are depicted details compatible with the stated truth conditions?
  4. Rights and disclosure: are key inputs and planned uses within the permissions and current terms the team relies on?
  5. Delivery/provenance: does the export work in context, and what origin/edit information can accurately be communicated?

A failure at gate two should usually stop polishing at gate five. This sequencing saves time.

C2PA-based Content Credentials may be part of gate five when supported, but they are not a substitute for gate three or four. Likewise, a legal review does not catch bad anatomy or a mobile crop. Each gate answers a different question.

Revision phase: change one class of error at a time

Revision should begin with a named error class, not a new mood word.

Examples:

  • crop failure;
  • identity drift;
  • perspective/contact error;
  • material inconsistency;
  • false product detail;
  • cultural-context concern;
  • unsupported claim;
  • provenance gap.

For each revision, write the smallest intervention likely to fix the class. If the intervention breaks a higher-priority anchor, revert it.

This creates a useful stop rule. When three consecutive revisions fix one issue while breaking another equally important issue, compare the cost of switching techniques. Manual paint, vector redraw, 3D, a real photo or a fresh concept may now be cheaper.

A workflow without a stop rule can consume unlimited generation time because every near miss looks like it is “one prompt away.”

Evidence and exceptions: copyright, tools and version drift

The U.S. Copyright Office's current AI materials are useful for one narrow question: in the United States, copyrightability turns on human-authored expressive material, and prompts alone do not automatically establish it. That is not a universal answer for every jurisdiction or every rights issue.

Tool capabilities are equally conditional. Reference-image controls, content credentials, model behavior and commercial terms can change. A workflow document should therefore record which product/version or current documentation was relied on, rather than turning today's interface into a timeless rule.

An exception also matters: very low-risk internal ideation may not need the same record as a paid international campaign. Scale the process to risk. Do not use “we are only brainstorming” as an excuse to upload confidential or unauthorized material, but do not burden every thumbnail with enterprise bureaucracy either.

Good governance is proportional, documented and revisable.

A one-page reusable template

A compact project sheet can carry the whole workflow:

Asset purpose: where it will appear and what action it supports.
Truth condition: illustrative, concept, factual product, reconstruction or documentary.
Non-negotiable anchors: three to five items.
Reference roles: identity / structure / style / material / constraint.
Current pass: composition, identity, geometry, material, detail or delivery.
Rejection tests: what makes a version unusable.
Rights notes: permissions, confidential inputs, current terms requiring review.
Human edits: layout, masks, compositing, paint, typography, selection.
Provenance: what the toolchain records and what it does not prove.
Stop rule: when to change technique.
Approver: who can actually close the asset.

That is enough structure for a small team without pretending the creative process can be fully automated. The workflow works because it turns hidden assumptions into inspectable decisions.

Research evidence before reference collection

The research folder should not be a mood-board dump. Split it into facts and visual inspiration.

Facts are materials that constrain truth: approved product dimensions, original character sheets, legal names, historical references, territory-specific campaign requirements, mandatory warnings and source permissions. Visual inspiration explores possibilities: lighting, rhythm, palette, framing, texture.

Mixing the two causes expensive mistakes. A mood image gets treated as a product specification; an obsolete campaign gets copied because nobody marked it archival; an AI-generated reference becomes the evidence for another AI-generated detail.

Give factual sources an owner and date. If a factual requirement changes, the owner triggers a review of dependent images. Inspiration can remain looser because its job is to provoke options, not certify facts.

Review the workflow itself after delivery

After the asset ships, spend ten minutes on a process review. Which gate caught the most expensive issue? Which reference created drift? Which manual edit saved the most time? Which documentation was never used?

Keep only lessons that change future behavior. “Use better prompts” is too vague. “For bilingual web heroes, block text zones before identity refinement” is reusable. “Do not use the archived 2025 product render as a dimensional source” is actionable.

This post-delivery review is how the workflow becomes genuinely repeatable rather than merely documented. It turns one project's surprises into the next project's defaults.

Do not freeze the workflow permanently. New model behavior, new product terms, new provenance standards and new campaign risks will change what deserves a gate. The stable principle is to make those changes explicit and evidence-based.

Approval is a decision, not a feeling

Close the workflow with an explicit approval record: who approved the asset, for which use, on which date, and with which unresolved limitations. A version approved for an internal pitch is not automatically approved for a paid ad. This small line keeps the next distribution decision from inheriting authority it never received.

A risk register keeps creative speed from becoming blind speed

Maintain a short risk register with five fields: risk, likelihood, consequence, owner and next evidence. Examples include “reference-image permission unclear,” “mobile crop not tested,” “product render may imply a non-final feature,” or “current tool terms need review before paid campaign use.”

The register should be small enough to read during a stand-up. Close risks when evidence arrives; do not preserve them forever for compliance theater.

This makes creative speed safer because the team can continue on low-risk work while a specific uncertainty is being resolved, instead of either freezing everything or pretending the uncertainty does not exist.

Archive the source of every current rule

For each high-impact rule in the workflow, record where it came from: current product documentation, an internal design decision, client policy, legal guidance or a past failure. This prevents folklore from hardening into process.

If a tool changes how reference controls work, only the rules dependent on that documentation need review. If a client relaxes a crop requirement, the production template can change without reopening copyright analysis.

This source-of-rule discipline makes the workflow easier to maintain because the team can update a dependency rather than rewrite the entire system.

It also exposes statements that have no authority at all—exactly the kind of “we always do it this way” habit that tends to survive long after its original reason disappeared.

Sources

Related Reading