Pikes

Build a brand world: a framework for exploring AI creative in CPG

A framework and step-by-step workflow for exploring creative directions with AI, with a sample project you can open and copy.

TL;DR

  • AI made creative volume cheap. The hard part now is exploring lots of directions without ending up with what looks like five different brands.
  • The fix is a brand world: lock the pack, anchor every image to one theme, and vary the style on purpose.
  • We walk through it with a flavor world for Pikes Cherry Blossom, a fictional soda: 1 can, 9 reference styles, two rounds with the Pikes agent, six finalists.
  • Steal the 8-step workflow and the copy-paste agent brief below.
Open the Pikes Cherry Blossom canvas
Pikes Cherry Blossom can beside a glass of pink soda with cherry blossomsPikes Cherry Blossom can on a pink cloth by a pool with cherriesPikes Cherry Blossom can among piled fresh cherries under blue sky
Three of the six finalists from the Cherry Blossom flavor world. Same can, same world, different styles.

AI has made exploring creative directions almost free. A brand team can try nine looks for a new flavor in an afternoon, where a photoshoot might cover one or two.

Most teams explore badly, in one of two ways. Some prompt at random and end up with a pile of images that look like five different brands, full of the model's defaults: the same porcelain vase, the same mid-century kitchen, golden hour on everything. Others lock one look too early and get 200 versions of the same shot. Neither gives a creative director real options to choose from.

A brand world sits in between. You fix two things: the pack, which has to be exactly right in every image, and the world, a theme that belongs to this product. For a drink, that's usually the flavor: its ingredients, colors, textures and mood. Then you vary the style on purpose, across very different references. The styles give you range. The world makes that range read as one brand.

As with most AI creative, quality comes from what you feed the model far more than from how cleverly you word the prompt.

The brand world framework

Every image in a brand world is built from three layers. Two stay fixed. One moves.

Fixed1. Pack truth

The exact pack from every angle, its real size, and what comes out of it. It never gets redrawn. If the label is wrong, nothing else about the image matters.

One theme2. The world

The sensory story the brand owns. For a beverage, the flavor: hero ingredients, the liquid, a palette that works with the pack. For skincare, scent and texture. For laundry, freshness.

Varied on purpose3. Style

The look of each shot: composition, set, light, camera, energy. This is where you want range.

The world is what makes nine different styles feel like one brand. The pack truth is what makes each of them usable. What you're after is a shortlist of real directions to guide the AI agents for world building.

Field study: a flavor world for Pikes Cherry Blossom

Pikes Cherry Blossom is a fictional soda we made up for this guide. We wanted to explore its flavor world and bring the taste of the drink to life. It runs in eight steps, and the four pages of the project canvas follow the same order, so keep it open in another tab.

01

Lock the pack truth

Goal: give the model everything it needs to reproduce the pack exactly, in any scene.

What to do: photograph every side of the pack in high resolution, with good light. Add a scale reference, ideally the product in someone's hand. (No in-hand photo? Generate one and have someone confirm the size reads right before you go further.) Then add a contents reference. For a drink, that's the real liquid color. Without it, “cherry soda” becomes whatever red liquid the model imagines.

Hand holding the Cherry Blossom can for scaleCherry Blossom can render at an angleFresh cherries as the flavor referenceGlass of the real pink-red liquid

The inputs: scale in hand, the can, the flavor, the real liquid color.

Common mistake: using low-resolution AI-generated images as your source. Hidden color artifacts compound with every regeneration. Use a real photo if possible, a 3D render if not. Even a smartphone photo performs well.
02

Define the world

Goal: decide what every image has to express before you look at a single style reference.

What to do: write a short brief, one line each: hero ingredients, what the product looks like out of the pack, a palette pulled from the pack itself, textures and props, the mood in three words, the things that would break it (your forbidden invents), and the claims printed on the pack that must stay exact.

In the sample: cherries and cherry blossoms. The real liquid color. Pinks from the can, with white and the flower's yellow as accents. Petals, rippling pink water, condensation, a regular glass.

Common mistake: choosing a palette that looks great on its own and fights the can. That's why the sample brief makes color harmony with the packaging an explicit rule.
03

Build a style board

Goal: go wide on purpose.

What to do: pick 6 to 9 references that are deliberately different from each other. Cover a range of camera angles (overhead, eye level, macro), sets (studio, tabletop, water, outdoors), light (soft, hard, backlit) and energy (a still life vs. a pour or a hand in frame). Pull them from Pinterest, editorial, competitor ads and your own past shoots, for internal exploration only. Then save the board. You'll reuse it for the next flavor.

In the sample: nine references that all bring a flavor to life in different ways, from pouring the liquid to a can sitting on ice among its ingredients.

Common mistake: using AI-looking references, or references with no real range in style. A good board also crowds out the model's defaults. Give it something better to copy and the porcelain, the kitchen and the golden hour disappear.
04

Brief the agent

Goal: one brief that carries every input, every rule and every setting.

What to do: select everything on the canvas (pack renders, flavor and liquid references, the scale shot and the style board) and send one message. Here's the exact brief from the sample:

Looking to explore some product brand universe directions for this Pikes cherry blossom drink. Attached are the can renders from different angles, the cherry image, and the scale of the can for reference. I've also attached 9 reference styles I like. I'd like your help creating two concepts based on each of the 9 references. Please make sure to localize the imagery so it reflects the flavor profile of the drink, and that the color composition is harmonious with the packaging design. I'd like to generate these using the Consistency Pro model at 2K resolution and 1x1 dimensions. Before generating any assets, please ask me any questions you need for alignment.

It names every input, sets a number of concepts per reference, ties every image to the flavor, sets a color rule relative to the pack, and fixes the model, resolution and ratio. The last line does the most work: the agent comes back with questions before it spends a single credit.

Here's a fill-in version for your own product:

Looking to explore [WORLD] directions for [PRODUCT]. Attached are the pack renders from [N] angles, [INGREDIENT AND CONTENTS IMAGES], a scale reference, and [N] reference styles I like. Create [2] concepts for each reference. Localize every image to the [WORLD]: [HERO INGREDIENTS, TEXTURES]. Keep the color composition harmonious with the packaging: [PALETTE]. Keep the pack exactly as shown. Don't show [FORBIDDEN INVENTS]. Generate with [MODEL] at [RESOLUTION], [RATIO]. Before generating anything, ask me any questions you need for alignment.
On models: we used Consistency Pro at 2K because the pack has to stay readable, and this model is built for product detail consistency. Creativity Pro (4K) could also work and tends to be stronger on photorealism, but it holds text less well. See which Pikes model to use.
05

Generate initial concepts

Goal: see what each style can do with your world.

What to do: ask for a set number of concepts per reference. One is too few to judge a style, and many more flood your review. Generate 2 to 4 outputs per run and pick the best. Choosing costs less than endless re-prompting.

Round one: can with cherry blossoms and ice on redRound one: tall glass of dark liquid with a canRound one: can floating on rippling pink waterRound one: dark glass and can with cherries on black

Round one. Strong directions, but the liquid color drifts where the liquid reference wasn't attached.

Common mistake: judging round one on polish. You're looking for directions that carry the flavor. Polish comes later.
06

Fix only what failed

Goal: tighten the world without starting over.

What to do: write one short note that adds whatever was missing and names whatever is banned. Redo only the misses. In the sample, round one missed on the liquid color. Here's the note we sent:

For the ones where liquid color is shown, make sure the liquid input image is also included. Redo those ones only. Also do not show the martini like glasses since this is a pop so a regular glass makes more sense.
After the fix: glass of pink soda and can with cherriesAfter the fix: glass and can on a yellow and pink gradientAfter the fix: pouring the can into a glass with a strawAfter the fix: glass and can with cherry blossoms

After the note: the real liquid color, in a regular glass.

That note is the whole framework in miniature. The liquid shots were missing their contents reference, so it adds one. And it redoes only what failed, which saves credits and leaves the good concepts alone.

Other defects we see often: bottle necks come out transparent (call it out and regenerate). Bottle bases melt into the surface (add an explicit constraint; it usually resolves within 2 to 3 regenerations). Hard colors like teal drift (fix the source image or clean up in post, because prompting rarely works).
07

Review and shortlist

Goal: pick the directions worth producing, without letting a wrong pack slip through.

What to do: review in two passes. Truth first: is the pack exact, are the claims intact, is the liquid right, are there any forbidden invents? That pass is pass or fail. Then taste: which directions go forward.

Finalist: can in a pile of cherriesFinalist: can on a pink cloth by a poolFinalist: can with a cherry on top among ice and cherriesFinalist: can with cherries and ice on redFinalist: glass and can with cherry blossomsFinalist: can surrounded by cherries under blue sky

The six finalists, tagged and commented on right on the canvas.

Every rejection became a rule, too. The martini glass is now banned from every future Cherry Blossom brief.

08

Produce from the shortlist

Goal: turn finalists into ads without breaking the pack.

What to do: turn a finalist into an editable frame and put the headline on its own layer. Expand to 1:1, 4:5 and 9:16 and check the Meta safe zone. Budget a Photoshop pass for small label text. Then send everything to Figma, a board or another project.

In the sample: the editorial can shot becomes a “New Flavor Drop!” ad, then gets resized and exported to a Pikes board or Figma.

Parting thoughts

Flavor is one kind of world. For supplements, the world is the gummy or powder flavor and the key ingredients. For personal care, it's scent notes and formula texture. For home and laundry, it's scent and freshness.

AI didn't make creative direction less important. It made it most of the job. The model will give you as many images as you ask for. It won't decide what stays fixed and what gets to move. Taste should not be outsourced to AI.

Questions

What is a brand world in AI creative?

A setup with three layers: the pack, which stays exact in every image; the world, one theme the product owns, such as its flavor; and the style, which you vary on purpose. The world is what makes very different styles read as one brand.

How many style references should I use?

6 to 9, chosen to be deliberately different in camera angle, set, light and energy. Ask for 2 concepts per reference so you can judge each style.

Which Pikes model should I use to explore a brand world?

Consistency Pro at 2K when the pack has to stay readable. Creativity Pro at 4K is stronger on photorealism but holds label text less well.

Copy the setup for your brand

Open the canvas to see the whole exploration, then run the same eight steps on your product. Questions: leo@pikes.ai