Last verified 23 September 2026. Gemini's standard image model takes several reference images on every plan. Adobe calls only its own Firefly models commercially safe, not partner models in the same app. Midjourney images are public below its $60-a-month Pro plan. Details for experts at the end of this book.

The idea in one page

A look is a list of decisions. "Clean, modern, professional" is a mood, and every generator reads it differently every time. What keeps a set together is five things you write down once:

References do what words can't. A picture of the palette and line style tells the generator more than a paragraph. But each reference needs one job ("colours and shapes", "spacing", "line and texture"), and a source you can name. The best references are ones you made yourself: three simple shapes in your template's colours, drawn in PowerPoint in five minutes. A picture from a web search can't go into the kit, however well it fits. You don't know who owns it, and the generator may copy it closely.

Half of consistency isn't in the generator. The fonts, the crop, the background, the label and the position on the slide come from your template. If the template is shared and everyone uses it, the images can vary a little and the set still looks like one set. If everyone crops and captions by hand, even matching images drift apart.

Only one line changes. The prompt frame is the four-part brief from T09-L01 · Make One Image or Narrated Clip, and Label It (subject, style, composition, exclusions), with the style, composition and exclusions frozen. Each colleague changes the subject line and nothing else. When a result drifts, fix the frame for everyone, not the prompt for yourself.

Test it on someone else. A kit that only works for its author is a note to self. The real test: a colleague who wasn't involved makes one new image from the folder alone, and a third person can't tell which image is the colleague's.

Rights and people, for the whole set. Three questions from T09-L01 · Make One Image or Narrated Clip, and Label It still hold: may you upload each reference, does the plan allow your use, and do you have consent from any recognisable person? At Level 2 there's one more: the answers live in the kit, so the next colleague doesn't have to ask them again. If the set may never show a real person, write that into the exclusions.

Worked example: two kits, four quadrants

The method comes from T01-L01 · From Idea to Working Result: The Four-Quadrant Method. Here it is for a research group, then for a company.

Lab: one look for the group's project slides

Structure it. Aylin, a postdoc, takes on the job after the retreat. "Four people make illustrations for their project slides, the group website and the site visit. A useful result: a kit in the group drive that lets anyone make an illustration in about ten minutes that matches the others, and nobody mistakes it for data."

Understand it. She collects the eight existing images and lays them out on one slide. She lists what differs: medium (3D, watercolour, flat), colours, shadows, and whether there's lettering. Then she tries the obvious fix. She writes one long style prompt and asks Gemini for all four students' subjects. The results are closer, but the blues still vary and two images have a drop shadow she didn't ask for. She tries attaching her favourite of the old images as a reference. That's better, and the next image copies its layout too, with the same round cell in the same corner.

That changes her problem. The words were never going to fix the colours, and one reference does too many jobs at once. She also notices that the worst clash on the retreat slide came from four different crops and two fonts, and the generator had nothing to do with that. Back to Structure it: "A kit with three purpose-made references, a frozen prompt frame, and the group template doing the crop, font and label. Illustrations only: abstract cells, tubes and arrows, never anything that looks like a micrograph or a result."

Choose tools. Gemini, because all four students already have it, it takes several reference images at once, and its visible watermark keeps the images honest as illustrations. The alternative she considers is Midjourney, which has a strong style-reference feature. She rejects it: on the cheaper plans the images are public, and the group doesn't want unpublished project topics in a public gallery. The group's PowerPoint template does the rest.

Build it. In PowerPoint she draws three reference images from plain shapes in the template colours: REF-A, a teal rounded rectangle and an orange circle on off-white, for palette and shape; REF-B, the same shapes pushed to the left two-thirds of a wide frame, for spacing; REF-C, one shape with a thin dark outline and no shadow, for line. She saves each one as a picture file (right-click the group of shapes, Save as Picture, PNG). Her style sheet fits on one page: the three hex codes, "flat vector, rounded, thin outlines, no shadow", the exclusions, and the crop sizes. The prompt frame:

She makes four images herself, one per student topic. For each she opens Gemini, clicks Add files in the prompt box, then Upload files, picks the three reference PNGs and pastes the frame with the new subject line. She rejects one with a faint scale bar in the corner, because a scale bar suggests real data. She places them in the template, which sets the crop and adds the line "Illustration generated with Gemini from the group's style kit."

Then the real test. Jonas, a new master's student, gets the folder link and a subject: "a pipette tip releasing a drop into a well". No chat with Aylin. His first image has a gradient background. He reads the style sheet again, sees no gradients is in the exclusions but was missing from his pasted frame, and pastes it whole. The second matches. At the next group meeting the PI looks at five slides and guesses wrong about which one Jonas made. Aylin adds one line to the kit: "Paste the whole frame; don't retype it."

What it taught her: the references and the template carried most of the look. The prompt mostly had to stay out of the way.

Company: pictures for a three-author onboarding guide

Structure it. Marek in HR owns the onboarding guide. "Three authors, twelve pictures, one guide that new employees read on their first day and that goes on our careers page. A useful result: pictures that look like one guide, that we may use publicly, and that show no real employee."

Understand it. He asks the authors how they made their pictures. The IT author used Gemini with no reference. The facilities author uploaded a colleague's photo from the intranet to "get the office look", and the result looks like her. The HR pictures came from a free stock site with a licence nobody saved. Then he checks the careers page plan: it's public and commercial, a recruiting page for the company. Back to Structure it: "One kit, references we made or licensed, no real person in any image, and a generator whose commercial terms our legal contact can read in five minutes."

Choose tools. Adobe Firefly, with one of Adobe's own Firefly image models. The same app also offers partner models from other companies, which Adobe doesn't vouch for in the same way. Adobe states its own models are commercially safe, trained on licensed pictures and public-domain material (that nobody holds rights to any more), and the company already pays for Creative Cloud, Adobe's subscription. The alternative is Gemini with several references, which the IT author prefers. Marek keeps Gemini for internal drafts and Firefly for anything on the careers page.

Firefly has a Style reference slot, where you upload one image whose look the new picture should follow. Marek draws that reference himself from the brand colours and simple shapes in the company's Word template, so its source is never in question. The colleague's photo is deleted from the kit, and the exclusions now say no people, no faces, no hands. People in the guide are shown as simple silhouettes drawn in the template.

Build it. The kit goes into a Teams channel: the style sheet, the reference with a line saying who made it and when, the prompt frame, the guide's Word template, and a label line: "Illustrations generated with Adobe Firefly." Each author regenerates their pictures with the frame, with Marek's reference in the Style reference slot. The facilities author's first attempt at "a desk with a laptop and a plant" includes a fake logo on the laptop, so "logos, brand marks" goes into the exclusions for everyone. Then the test: a colleague from finance, who wrote nothing, makes the thirteenth picture for a new chapter on expense claims. Marek shows all thirteen to the office manager, who can't pick out the new one.

What it taught him: the rights question took longer than the style. Once the references had a named source and people were excluded, the rest was routine.

In both cases the kit made the style portable. The colleague test showed where it still leaked.

What goes wrong

Adjectives instead of references. Symptom: every prompt says "modern, clean" and the colours still drift. Fix: make three references in your template's colours, give each one job, and attach them every time.

One reference doing everything. Symptom: every new image copies the layout of the reference, not only its look. Fix: split it into separate references for colour, spacing and line, and tell the generator not to copy the arrangement.

A reference without a source. Symptom: nobody knows where a reference came from, and the result looks just like someone's photo or artwork. Fix: keep only references you made or licensed, and write the source next to each file. Delete anything you can't name.

A real person as "just a reference". Symptom: a generated face looks like a colleague. Fix: exclude people in the kit, or use written consent for this exact use. An illustration of a person who looks real needs the same care as a photo.

The kit lives in one chat. Symptom: only the author can make a matching image, from their own history. Fix: keep the frame, references and style sheet in the shared folder, and have colleagues paste the frame whole.

Mixing generators halfway. Symptom: the last three images have different edges and colours. Fix: name one generator and model in the kit. If you have to switch, remake two old images with the new one and compare before going on.

Fixing the prompt for yourself. Symptom: each person tweaks the frame, and the set drifts again within a month. Fix: one person maintains the kit. Changes go into the shared frame, with a date.

Do it yourself (60 minutes)

Use a real set: a series of slides, a website section, a guide, a poster.

1. Structure it (10 min). Write what the set is for, who sees it, how many images, and what they must never suggest (real data, a real person, a product feature).
2. Understand it (10 min). Lay out the images your team has already made on one page. List what differs. Try one image with the generator using words only, and note what drifts.
3. Choose tools (5 min). Pick one generator and model for the set, with one line on why and whether your use is commercial. Name the shared folder and template.
4. Build it (25 min). Draw three references in your template's colours and write their sources. Write the one-page style sheet and the prompt frame. Make four images with the three references attached to every request, reject what breaks the style sheet, and finish them in the template with the label line.
5. Test it (10 min). Send the folder to a colleague with one subject line and nothing else. Put their image with your four and ask a third person to pick it out. Fix the kit where your colleague got stuck.

Exit check

What you hand in: a shared style-kit folder and the set it produced:

• the style sheet (one page): palette as hex codes, shapes and medium, composition, exclusions, crop sizes;
• three references, each with its job and its source;
• the prompt frame, and the generator and model it's for;
• the template and the label line;
• five finished images, one of them made by a colleague from the folder alone, with a note on what they had to ask or fix.

It passes when a third person can't pick out the colleague's image, and every reference in the folder has a recorded source.

Rule to remember

A look belongs to the team when a colleague can make it without you.

Details for experts

Checked on 23 September 2026. You don't need this to follow the book; it's for setting things up or answering a colleague.

Further reading and tools

Next books

• T09-L01 · Make One Image or Narrated Clip, and Label It: the four-part prompt and the AI label this book builds on.
• T09-L03 · Build a Caption Pipeline Your Team Can Rerun Every Week: when the set becomes a batch that runs every week.
• T12-L02 · Clean the Copy, Label the Result, Spot the Trap: what may be uploaded, and how to label it.
• T02-L03 · Build a Shared Assistant Your Team Can Test and Take Over: shared instructions for text, the same idea as a shared kit.

Tools

• Gemini: Taught. Several reference images per request, and a watermark on every image.
• Adobe Firefly: Taught. Style reference with commercially safe Firefly models, for public and commercial sets.
• Midjourney: Catalogued. Strong style references and style codes; public by default below the Pro plan.
• Nano Banana Pro: Catalogued. Gemini's larger image model on paid Google AI plans.
• C2PA: Catalogued. The standard behind content credentials, a record inside an exported image that says how it was made.