Skip to content

AI Image Generation and Design · Prompting for images · lesson 5 of 18 · 7 min

Styles, references and responsible style direction

Style is where brands are won or lost

Two campaigns can show the same subject and feel completely different because of style: lighting, palette, texture, rendering technique, lens and composition. For brand work, style consistency matters more than any single stunning image.

Describing style in words

Break style into components rather than relying on one label:

Style spec
Medium:        editorial photograph / flat vector / 3D render / gouache
Era/influence: 1970s print ads, Bauhaus posters, contemporary minimal
Lighting:      soft diffused daylight, hard flash, studio rim light
Palette:       muted earth tones; teal + sand; high-contrast monochrome
Texture:       film grain, paper texture, glossy plastic, matte clay
Detail level:  simple shapes, highly detailed, painterly brushwork
Composition:   centered, generous negative space, low angle

This "style spec" becomes part of your campaign prompt library.

Reference images

Many tools accept images to guide results. Common types:

| Reference type | What it controls | Example use | |---|---|---| | Style reference | Colors, textures, rendering | Keep a campaign look consistent | | Subject/character reference | A person, mascot or product's appearance | Same mascot across ten posts | | Composition/structure reference | Layout, pose, depth, edges | Match a planned layout or pose | | Image prompt / remix | Overall resemblance | Variations of an existing concept |

Most tools let you adjust reference strength. Too weak and the style drifts; too strong and outputs become near-copies with little variation.

Rights in references

You must have the right to use images you upload as references. Safe sources:

  • Your own photographs and artwork.
  • Assets owned by the client (with their permission).
  • Images you have licensed for that purpose (check the license permits this use).
  • Your previously approved AI outputs for the campaign.

Using someone else's copyrighted artwork as a reference to produce closely similar work risks infringement and is ethically problematic, even if the tool technically allows it.

Artist names and "in the style of"

Prompting "in the style of [living artist]" raises ethical and legal concerns:

  • Many artists object to their distinctive styles being replicated without consent or compensation.
  • Outputs that closely imitate specific works may infringe copyright; style in the abstract is generally not protected by copyright in many jurisdictions, but close imitation can cross into copying, passing off or unfair competition depending on the facts and the country.
  • Some platforms restrict or discourage prompts naming living artists.
  • For client work, it creates reputational risk.

Better practice: describe the visual characteristics you want (medium, era, palette, texture, composition) or commission the artist directly. Historic art movements ("Art Deco poster style", "ukiyo-e inspired") are generally safer as descriptors than individual living artists.

Brand and trademark cues

Avoid prompting for other brands' logos, packaging or characters. Outputs can include recognizable trademarks or trade dress, which creates legal and platform-policy risk. Use generic descriptions and add your own brand assets in the design tool.

Creating an ownable style

For a brand or creator, develop a distinctive AI-assisted style that is yours:

  1. Start from your brand strategy and personality sliders.
  2. Build a style spec in words.
  3. Generate candidates and select a small set of approved "hero" images.
  4. Use those approved images as style references going forward.
  5. Document the prompt snippets, reference images and strength settings.
  6. Review periodically — model updates can shift results.

Worked example: an illustrated style for a finance education creator

  • Personality: calm, smart, approachable.
  • Style spec: flat vector illustration, soft rounded shapes, limited palette of navy, mint and warm sand, subtle paper grain, simple characters without facial detail, generous white space.
  • References: three approved illustrations from the first batch, used as style references at medium strength.
  • Result: consistent illustrations for carousel covers across months, without imitating any particular artist.

Common mistakes

  • Relying on a single style keyword that the model interprets inconsistently.
  • Uploading references you have no rights to.
  • Prompting with living artists' names for commercial work.
  • Setting reference strength so high that outputs are near-duplicates.

Brand-consistent generation in 2026: three levels

| Level | What you provide | Examples of features (check current availability) | Best for | |---|---|---|---| | 1. Written style spec | The style spec above, reused in every prompt | Any tool | Quick consistency, low setup | | 2. Style and subject references | 1–10+ approved images per style or subject | Style references in Midjourney and Ideogram; Firefly style and composition references; Recraft custom styles built from reference images; multi-image references in Gemini image models | Campaigns, illustration series, recurring characters | | 3. Custom models | A curated, rights-cleared training set | Adobe Firefly Custom Models (plan-dependent); fine-tuning open-weight models (for example LoRA adapters for FLUX or Stable Diffusion) | High-volume brands that need the same look for months |

Recraft, for example, describes its V4 styles as built from one to ten reference images with no training step, and Google describes Gemini 3 Pro Image as accepting up to 14 reference images. Treat these as vendor claims and test them on your own brand.

Hands-on: build a brand style kit for generation

BRAND STYLE KIT — [brand] — v1.0 — owner: [name]
1. Style spec (words):   [medium, era, lighting, palette, texture, detail, composition]
2. Style references:     style_ref_01..06.png   (rights: own/client-owned/licensed)
3. Subject references:   product_front.png, product_34.png, mascot_sheet.png
4. Prompt snippet:       "[20-40 word style snippet pasted into every prompt]"
5. Tool settings:        [tool/model], reference strength: medium, aspect ratios used
6. Do / don't gallery:   approved_01..04, rejected_01..04 (with one-line reasons)
7. Review:               brand owner approves new references; kit reviewed quarterly

Steps:

  1. Select six style references that share a look but vary in subject. Too-similar references produce near-copies.
  2. Generate twelve test images across three subjects using the kit.
  3. Lay them out on one board (Figma, Canva or a moodboard tool) and check: does the palette hold? Does the texture hold? Would a customer recognize the brand without the logo?
  4. Adjust reference strength or swap weak references, then freeze the kit as v1.0.

Before/after: references versus words alone

| | Before: words only | After: words + six style references | |---|---|---| | Palette across 12 images | Drifts between warm and cool | Holds to the brand's teal and sand | | Line and texture | Mix of flat, 3D and painterly | Consistent flat shapes with paper grain | | Recognizable without logo | Rarely | Usually |

(Illustrative outcome; test on your brand.)

Summary

Describe style through components, use references you have rights to at an appropriate strength, avoid imitating living artists and other brands, and build an ownable, documented style from approved hero images.

Video lecture: Styles, references and responsible style direction

Lecture coming soon · 10 chapters · about 8 minutes. Read the full transcript below.

  1. Styles and references
  2. Why style matters
  3. Three levels of consistency
  4. Real features (check current docs)
  5. Rights in references
  6. Worked example 1: a finance creator's covers
  7. Worked example 2: an agency style kit
  8. Watch me do it: a 30-minute style kit
  9. Common mistakes
  10. Recap and try this now

Lecture transcript

Styles and references

Scroll through any feed and some brands jump out even before you see their logo. You just know it's them. That's style doing its job. Now imagine producing fifty AI images for a campaign, and every one looks like it came from a different agency. That's the default outcome if you rely on luck. In this lecture, you'll learn how to describe style in words, how to use style and subject references properly, when a custom model makes sense, and how to stay on the right side of artists' rights. By the end, you'll be able to build a brand style kit that makes your AI images recognizable without a logo.

Why style matters

Why does style matter more than any single great image? Because brands are built on repetition. People recognize patterns, not one-offs. If your palette, lighting and texture stay consistent, every image adds to the same memory. If they drift, every image starts from zero. Think of a band. You recognize your favorite band from the first few seconds of a new song, because the sound is consistent even when the song is new. Style is your brand's sound. And AI makes drift incredibly easy, because each generation pulls toward the model's defaults. So consistency doesn't happen by accident. You have to engineer it.

Three levels of consistency

Level one is words. Break style into components instead of one vague label. Medium, like flat vector or editorial photo. Era or influence, like nineteen-seventies print ads or Bauhaus posters. Lighting. Palette, described in words. Texture, like film grain or paper grain. Detail level. And composition habits, like generous negative space. That's your style spec. Level two is references. You upload approved images that show the style, or the subject, like a product or a mascot. Most tools let you set how strongly the reference applies. Too weak, and the style drifts. Too strong, and you get near-copies. Level three is a custom model, trained on a curated, rights-cleared set. Powerful, but only worth it at high volume.

Real features (check current docs)

Let's name some real features, and remember to check current availability, because this changes fast. Midjourney and Ideogram offer style references. Adobe Firefly offers style and composition references, and Firefly Custom Models on some plans. Recraft builds custom styles from a handful of reference images without a training step. Google's Gemini image models accept multiple reference images, and Google describes its Pro image model as handling up to fourteen of them. And if you work with open-weight models like FLUX, teams fine-tune small adapters, often called LoRA, on their own images. Here's the key idea. The tool matters less than the discipline. Curated references, a documented spec and a review step beat any single feature.

Rights in references

Now rights, because references are where people get into trouble. You must have the right to use every image you upload. Safe sources are your own photos and artwork, assets the client owns and has approved, images licensed for that use, and your own previously approved outputs. Using another artist's work to make close imitations is risky legally and ethically, even if the tool allows it. The same goes for typing in the style of, followed by a living artist's name. Many artists object, some platforms restrict it, and close imitation can cross into infringement or passing off depending on the country. Describe visual characteristics instead, or better, commission the artist.

Worked example 1: a finance creator's covers

Worked example one, simple. Priya is a finance education creator in London who publishes a weekly carousel. She wants a spot illustration on every cover. With words alone, her twelve test images drift. Some are flat, some are 3D, and the palette swings between warm and cool. So she writes a style spec: flat vector shapes, subtle paper grain, rounded corners, a teal and warm sand palette, one accent of coral, lots of empty space. Then she picks her six best approved images as style references at medium strength. The next twelve images look like a family. And here's the test she uses. Cover the logo. Would a follower still know it's her? Now the answer is yes.

Worked example 2: an agency style kit

Worked example two, a business scenario, with illustrative details. An agency in Karachi runs social content for a regional snack brand, about sixty assets a month across three markets. The brand has a cheerful mascot and a specific packet design. The agency builds a brand style kit. Style spec in words. Six style references. A mascot reference sheet from the brand's own illustrator, and front and three-quarter photos of each packet. Then they set a rule. Packets are never generated from text alone. They're always composited from real product photos. After a month, the agency reviews results. The mascot stays consistent with references, and the one time the model altered the packet, the review step caught it before posting.

Watch me do it: a 30-minute style kit

Watch me do it. I'm building a brand style kit in about thirty minutes. First, I open a board in Figma and drag in twenty candidate images from the brand's own library. I pick six that share palette, texture and light but show different subjects. That variety matters, or the model just copies one image. Next, I write a twenty to forty word style snippet that I'll paste into every prompt. Then I generate twelve test images across three subjects, and lay them out on the board. I check three things: palette, texture and recognizability without the logo. Two references are pulling too hard toward one composition, so I swap them. Finally, I save the kit as version one point zero, with the owner's name and a review date.

Common mistakes

Common mistakes. Using one label, like modern or clean, instead of a real style spec. Picking references that are all the same image, which produces copies instead of a style. Setting reference strength to maximum and losing all variety. Mixing references from different styles and expecting consistency. Uploading images you don't have rights to. And never updating the kit. Models change. A new version of your favorite tool may interpret the same references differently, so re-test your kit whenever the model updates, and keep your approved outputs as the real source of truth. Your kit is a living document, not a one-time setup.

Recap and try this now

Let's recap. Style is your brand's sound, and AI drifts toward defaults unless you engineer consistency. Use three levels: a written style spec, curated style and subject references at sensible strength, and custom models only when volume justifies them. Only upload images you have rights to, and describe characteristics instead of copying living artists. Try this now. Build version one of a brand style kit for yourself or a client using the template in the lesson. Six references, a style snippet, twelve test images, and one board. Then do the logo test. Cover the logo and ask, would someone still know it's us?

Key takeaways

  • Consistency comes from three levels: a written style spec, curated style/subject references, and (at high volume) custom models.
  • Choose references that share a style but vary in subject, and tune reference strength to avoid drift or near-copies.
  • Only upload images you have rights to; describe visual characteristics instead of imitating living artists.
  • Package the spec, references, snippet, settings and do/don't gallery as a versioned brand style kit.
  • Re-test the kit whenever a model updates; approved outputs are the source of truth.

Try it

Write a style spec for a brand, generate a batch of images, select three hero images, and document the prompt snippet and reference strength that reproduce the look consistently.