Consistent assets for your company or lab

2. Eight slides, eight visual languages

You ask an image generator for eight illustrations for next week's review. Each prompt describes the subject correctly, yet the set looks assembled from eight unrelated projects. One image is photographic, one resembles a paper cut-out, two use different blues, and every object casts a different kind of shadow. A teammate replaces three images, making the mismatch worse.

This is the Level 2 Power user blast radius: your images now appear together in material that teammates review, reuse, or present. One inconsistent, misleading, or uncleared image can weaken the whole poster, talk, or product page.

The problem is not a shortage of style adjectives. Nobody saved what must remain constant, which references were approved, or which settings produced the first acceptable image. There is also no record showing whether the reference files may be used for generation. A recognizable person has appeared in one candidate even though no person was requested.

You need a small production system: a fixed brief, three permitted references, a reusable prompt skeleton, a do-not list, controlled settings, and one review record. The same system must work for a fictional Lab poster and talk or a fictional Company product page without turning generated scenes into evidence.

3. After this you can

• Define the visual constants that make five assets read as one family.
• Build a reusable prompt and three-reference kit another teammate can follow.
• Control variation with fixed settings and seeds where the selected tool supports them.
• Check the applicable terms for the actual model, account, input, and publication use.
• Release a consistent set with provenance, consent, accessibility, and human review recorded.

4. Prerequisites

• T09-L01 · Images, voice and video for your own work.
• One organisation-approved image generator and one image or slide editor.
• Permission to read the provider's current terms and account or plan details.
• A folder for prompts, approved references, candidates, editable masters, exports, and the release record.
• About 45 minutes and the synthetic fixtures below, or equally safe public or explicitly approved material.

Do not upload unpublished figures, participant or patient images, customer files, employee photographs, logos, confidential product screens, licensed artwork, or found images unless an accountable owner has approved that specific input and synthetic use. Use non-identifying synthetic material for this exercise.

5. The idea in one page

Consistency is a list of decisions, not a mood. Save these six controls before making the set:

Use references to show visual decisions that words describe poorly. Reference A can anchor colour and shape, B spacing and composition, and C texture and line treatment. Do not use a reference merely because you like it. Record its origin, permission, permitted channels, and whether alteration or use as a generative input is allowed. A public image is not automatically a permitted reference.

A seed, when exposed by the tool, can make controlled reruns easier. It does not guarantee identical output across model updates, tools, or changed prompts, and it does not prove provenance. Save the model or mode, aspect ratio, reference settings, prompt, date, and selected output as well. If the tool has no seed control, keep the reference set and fixed prompt unchanged and generate only a small batch.

Create facts and exact text outside the generator. Generated pixels are a poor authority for measurements, labels, product functions, logos, citations, or findings. Add checked copy in an editor, retain an editable master, and inspect the exported files at their real display sizes.

Terms check before production

Current-product check — last verified 2026-09-04: Terms, models, and plans change. Inspect the primary terms for the exact service, account, inputs, and channel. Check output rights, reference restrictions, commercial or publication permission, plan conditions, provider reuse and retention, visibility defaults, and required disclosure. “Open weights,” “commercially safe,” paid access, or access through another platform is not permission. If the planned use is unclear, record BLOCKED and do not upload or publish.

Use one row in the release record for the decision: service route | exact model shown | account/plan | primary terms URL | checked date | input permission | output/channel permission | visibility and provider-use setting | owner | PASS/BLOCKED. A blank or assumed field is BLOCKED, not permission.

Do not choose by brand name:

Consent is separate from provider terms. Permission to photograph someone may not cover a style reference, synthetic variation, advertising, or publication. Record permission for the input, transformation, portrayal, channels, and duration, or use no recognizable person. Disclosure does not repair missing consent.

6. The worked example: one kit, two settings

The fictional Aster-9 Lab and Northstar Desk Company projects use the same production skeleton. Each needs five coordinated illustrations. No real people, organisations, products, findings, or source media are used.

Step 1: lock the boundary and asset roles

Lab framing: The Aster-9 set supports a practice poster and internal talk about a synthetic literature-review workflow. The five roles are poster opener, corpus collection, screening, synthesis, and talk divider. The illustrations may show abstract cards, paths, and containers. They must not resemble experimental evidence, real apparatus, journal branding, a published figure, or a scientific result.

Company framing: The Northstar Desk set supports a private product-page prototype. The five roles are page hero, request intake, sorting, handoff, and section divider. The illustrations may show abstract cards, paths, and containers. They must not resemble a real interface, customer testimonial, employee, certification, available feature, or performance result.

The factual copy is synthetic and added later. Aster-9 is a fictional training workflow. and Northstar Desk is a fictional product concept. appear near the relevant set. Neither sentence claims a measured benefit.

Step 2: create three permitted references

Make the reference board yourself in the editor, using only rectangles, circles, lines, and these invented colours:

Save the three references as one labelled board and as separate files. The board is not a mood board scraped from search results. Its provenance entry states creator, date, editor, source elements, permitted synthetic-reference use, and Consent: not applicable; no person or voice represented.

Step 3: save the reusable kit

Use this shared skeleton for either framing:

PROJECT: [Aster-9 / Northstar Desk]
ASSET ROLE: [one of the five approved roles]
COMMUNICATION JOB: Illustrate [bounded concept], not evidence of a real event,
result, person, workplace, product interface, or available feature.

SUBJECT: [2-4 permitted abstract objects and one relationship]
STYLE: Flat geometric editorial illustration; rounded modular shapes; thin navy
outlines; cream background; one coral focal accent; subtle paper grain; one
small offset shadow; restrained, calm, non-photographic.
COMPOSITION: [orientation]; subject within left two-thirds; generous negative
space; one focal point; simple silhouette readable at thumbnail size.
REFERENCES: REF-A for palette and shape; REF-B for spacing; REF-C for line and
texture. Do not copy their exact arrangement.
EXCLUDE: people, faces, hands, logos, lettering, numerals, badges, watermarks,
screenshots, realistic instruments, realistic products, results, charts,
certification marks, dramatic lighting, gradients, and extra decorative objects.
OUTPUT: illustration only. Add all verified copy, labels, and disclosure later.

The kit also saves the three reference files, colour values #F4EFE6, #17324D, and #E56B5D, the five asset roles, aspect ratios, selected service route and exact model label shown by the service, all exposed generation and reference-strength settings, editor/export settings, and the do-not list. Record unavailable controls as not exposed; do not silently omit them. If a seed is available, save the accepted starting seed and treat it as a rerun aid, not an identity lock.

Use one concrete pack structure so a teammate does not have to infer where evidence belongs:

T09-L02-consistency-pack.zip
  manifest.md
  kit/prompt-skeleton.txt
  kit/settings-and-terms.md
  kit/REF-A.png  kit/REF-B.png  kit/REF-C.png
  review/contact-sheet.jpg
  masters/ASSET-01-to-05-editable-files
  exports/ASSET-01.png ... ASSET-05.png
  rerun/test-output.png

In manifest.md, list the project boundary, five asset roles, variable values, source and candidate IDs, final filenames, SHA-256 hashes for the three references and five exports, alt text, disclosure, edit history, reviewer, channel, and release decision. The ZIP is the single delivered artifact; its entries are components and evidence, not separate submissions.

Step 4: generate bounded variations

For each role, change only ASSET ROLE, SUBJECT, COMMUNICATION JOB, and orientation. Keep every style line, reference, exclusion, service route, model label, and exposed setting fixed. Generate two candidates per role. Give candidates stable IDs such as ASSET-03-C1; record the complete rendered prompt rather than only the template.

For the Lab screening image, use three abstract document cards passing through a rounded gate; one coral card remains in focus. For the Company sorting image, use three abstract request cards moving into two rounded trays; one coral card remains in focus. The structure is parallel while the meaning remains appropriate to each setting.

Review the ten candidates as a contact sheet. Reject any candidate with accidental writing, an invented mark, a realistic person or instrument, an implied result, a product screen, a new palette colour, incompatible texture, or a different visual depth. Do not crop a rights or identity problem out of sight. Record the candidate ID and rejection reason.

Step 5: finish and release the set

Choose one candidate per role. In the editor, apply the exact output dimensions and add only approved copy. Add useful alt text based on each image's communication purpose, not a list of every shape. Place the accurate disclosure AI-generated illustrations assembled from an original synthetic reference kit; copy reviewed by the project owner. where it remains associated with the set.

The reviewer compares all five exports side by side and checks palette, line weight, depth, spacing, texture, disclosure, factual implications, input rights, non-identifying status, alt text, and final crops. Before the handoff, define the rerun pass rule: the teammate uses the unchanged ZIP contents and one recorded asset row; the service route, model label, rendered prompt, references, and exposed settings match the manifest; the raw rerun contains no excluded element; and it passes all five named style checks: only the cream, navy, and coral palette; rounded flat shapes; subject in the left two-thirds; thin navy outlines; and subtle grain with one small offset shadow. Pixel identity is neither expected nor required.

Keep the manifest even if the service supplies Content Credentials: metadata can be stripped and does not prove truth, rights, or consent.

7. What goes wrong

Adjectives replace references

Symptom: every prompt says “clean, modern, professional,” but colour, depth, and texture still drift.

Fix: create three permitted references with one visual job each. Save exact palette, composition, and exclusion rules.

The reference board has no rights record

Symptom: a teammate collects attractive web images, and nobody knows whether generative use or publication is permitted.

Fix: record origin and permission for every reference. Replace unclear material with original synthetic geometry or explicitly approved assets.

A seed is treated as a guarantee

Symptom: a model update or prompt edit changes the look even though the same seed was entered.

Fix: retain the complete kit and accepted outputs. Recheck consistency after any tool, model, reference, prompt, or setting change.

The team mixes tools halfway through

Symptom: the last two images have different edges, colour handling, and depth.

Fix: complete the set in one approved tool and mode. If migration is necessary, regenerate a small comparison set and approve a new kit version.

Consistency preserves a bad implication

Symptom: all five images consistently look like real experiments, customer evidence, or released product screens.

Fix: return to the communication boundary. Use unmistakably illustrative forms and remove unsupported visual claims; consistency cannot make fiction factual.

Consent is inferred from access

Symptom: a colleague's portrait is used because it was in a shared folder or earlier slide deck.

Fix: obtain specific documented permission for synthetic reference and release, or redesign with no recognizable person. Keep the decision in the record.

The kit works only for its author

Symptom: another teammate cannot reproduce the style because settings, references, or variable fields are missing.

Fix: hand the unchanged kit to that teammate for one controlled test-asset rerun. Keep the rerun output, settings record, and any missing-instruction note in the consistency pack; update the kit and rerun before release if the test fails.

8. Do it yourself: build the set in 45 minutes

Minutes 0-5: choose Lab or Company and name a fictional project. Define audience, channel, five asset roles, what the set communicates, and what it must never imply.

Minutes 5-12: create three simple synthetic reference images. Assign one to palette and shape, one to composition, and one to line and texture. Record provenance and the non-identifying consent decision.

Minutes 12-18: write the fixed prompt skeleton, exact colours, aspect ratios, do-not list, service route, exact model label, exposed settings, and seed if available. Complete the terms row for the exact account and intended channel. Pause the exercise rather than guessing if approval or a primary term cannot be checked.

Minutes 18-30: generate two candidates for each role. Change only the allowed fields and retain each rendered prompt.

Minutes 30-37: compare candidates side by side. Select one per role and record every rejection reason.

Minutes 37-42: finish the five images in an editor. Add approved copy, disclosure, alt text, and consistent dimensions; then export.

Minutes 42-45: give the saved kit to a teammate for one controlled rerun of a selected asset, without private explanation. Keep the raw rerun output, rendered prompt, reference hashes, settings, UTC date, teammate identifier, and five-check comparison result inside the pack. If the teammate is not available in this window, leave the decision BLOCKED and finish after the observed rerun; an explanation of what they would do does not pass. Approve only if the rerun follows the kit and the five final exports visibly form one family.

9. Exit check

Deliver exactly one artifact: T09-L02-consistency-pack.zip, containing the five final exports and the reusable kit, manifest, terms decision, review evidence, and teammate's controlled-rerun evidence defined above. Do not submit the images, manifest, or rerun as additional artifacts.

It passes when the archive opens; contains exactly five final exports; records PASS for the applicable terms and input permissions; and its hashes match the three references and five exports. A reviewer must identify the shared palette, shape, composition, line, and texture rules across all five assets. The retained rerun must match the recorded route, model label, rendered prompt, reference hashes, and exposed settings, contain no excluded element, and pass all five style checks without private explanation. Every input must have a permitted source, no image may imply unsupported evidence or identity, and each export must be traceable to its prompt, candidate, edits, alt text, disclosure, and named human approval.

10. Rule to remember

Consistency is a saved kit, not a lucky prompt.

11. Further reading & tools

• Taught: T09-L01 · Images, voice and video for your own work - choose the medium, bound facts and rights, and label the result.
• Taught: AI image generation - create a fact-bounded brief and retain asset, accessibility, provenance, and release evidence.
• Taught: T12-L01 · What you may and may not paste into AI at work - decide whether a reference may enter the selected service.
• Catalogued: Midjourney Terms of Service (opens in a new tab) - check current account, ownership, company, and input conditions before use.
• Catalogued: Black Forest Labs licensing (opens in a new tab) - identify the exact FLUX model or service and its applicable terms.
• Catalogued: Adobe Firefly FAQ (opens in a new tab) - review current commercial-use, training-source, and Content Credentials statements.
• Catalogued: Google Gemini API Additional Terms (opens in a new tab) - review the current terms for Google's image-capable Gemini services.
• Catalogued: C2PA specifications (opens in a new tab) - understand signed provenance assertions and their limits.
• Catalogued: W3C Web Content Accessibility Guidelines 2.2 (opens in a new tab) - check text alternatives for non-text content.
• Catalogued: Tools index - compare current image tools only after defining references, intended use, and review requirements.