← back to blog

Why AI Can't Count the Products on Your Display

November 12, 2026··9 min read

You asked for three shelves with five facings across each. The render came back with nine on the top shelf, six in the middle and something between seven and eight at the bottom, because the last pack melts into the side panel. You write "exactly five" and generate again. Now it's four. Then seven.

It isn't your prompt. Image models don't count, and on a display the number they get wrong is the one the brand is paying for.

I've worked in point-of-purchase for close to four years and I build AI POP Displays. This is the failure I'd explain first to anyone sending AI renders to a client, because it looks like a detail and it's the spec.

Why does the render show nine packs when you asked for five?

Because nothing inside an image model counts. It generates a picture that fits the request as a whole, and "five" is one signal among many, next to the shelf, the pack, the brand color, the store and the light. What the model learned from retail photography is that shelves look full. A full shelf is the plausible picture, so it fills the shelf. The number of packs comes out of the width of the shelf and the size the pack happened to be drawn at, not out of your brief.

This isn't a hunch. A 2025 paper with the blunt title "Text-to-Image Diffusion Models Cannot Count, and Prompt Refinement Cannot Help" built a counting benchmark, asked fifteen text-to-image models for anywhere between one and fifteen objects, and had people check the results. Its conclusion is that all the state-of-the-art models it tested fail to generate the correct number of objects, with accuracy dropping significantly as the number of objects increases. Another study, on optimizing object counts, notes that a model asked for a specific number, such as ten balls, usually draws a significantly larger quantity. That's your nine for five. A third paper, "Counting Hallucinations in Diffusion Models", gives the failure a name, counting hallucination, the generation of an incorrect number of instances or structured objects. The six-fingered hand is the case everybody has seen.

Search for help with this and you land in one of two places. One is forum threads where someone asked an image tool for three of something, got five, and the reply is that counting isn't a strength of AI image generators. The other, the moment the word shelf goes in, is software that counts facings in photos of real stores to check planogram compliance. That's computer vision reading a photograph, not a model drawing a concept. Neither is about a generated display that somebody has to quote.

Why doesn't rewording the prompt fix it?

Because the number was never a control. It's a word.

The same paper tested this directly. It took its counting prompts and rewrote them in several structured ways, breaking the number into smaller groups, adding a layout, giving positions, and found that these simple interventions generally do not improve counting accuracy. For most of the models it tested, the rewritten prompts did worse than the plain ones.

People end up putting the number in capitals, spelling it out, repeating it, and regenerating until a render happens to land on the right count. Then the client asks for a taller header, the image is generated again, and the count moves with everything else. A count you got by luck is a count you lose on the next revision.

Why is a shelf the hardest case?

Everything that makes counting hard sits on a display at once.

Identical objects in rows. A shelf is the same pack repeated. A CVPR 2025 paper on accurate counts, "Make It Count", puts the difficulty in one line. The model has to keep a separate identity for every instance of an object, even when several look identical or overlap. That describes a shelf.

Larger numbers. The paper found accuracy falls as the requested number rises. A shelf of five across and three deep is fifteen packs on one tier, and a floor unit has several tiers.

Depth and occlusion. The back rows are half hidden by the front ones. Whether a shelf is two deep or three is barely visible in the picture, so the model is free to change it.

Perspective. A three quarter view compresses the shelves furthest from the camera, so even a right count is hard to read.

Different SKUs. A range of four packs in different sizes has to be counted per variant. A model balancing a picture draws them the same apparent size.

And the model has no failure state. Nothing comes back saying the count is off. The render arrives at the same confidence with five packs or nine, which is why a wrong count usually survives internal review and gets found by the person who prices the unit.

Why does a wrong count cost more on a display?

Because facings are capacity. The number of packs a shelf carries is what the brand is buying in every store, what the retailer's buyer is giving space for, and what the manufacturer prices. Shelf width, shelf depth, how many shelves and what the base carries all follow from it.

A stray apple in a stock image costs nothing. A display concept with nine facings where five fit does one of two things. Either the client approves capacity the unit won't have, and that conversation happens after the quote. Or the structure grows to reach the number in the picture, and the price grows with it. I went through the other ways a general model drifts on product load in why generic image AI fails at POP displays.

If there's a render in your last deck with a shelf nobody counted, put the same brief through the free plan with the real packs attached and count the two side by side.

Where does the count get decided in AI POP Displays?

Not in a word inside a prompt. In the inputs the render is built from, and in the views you check it on.

The pack is real. Product photos go in their own slot and are loaded as the product, at scale, facing forward, with each SKU kept separate. A count of your client's real pack is a count of something specific. A count of a lookalike drawn at a guessed size means nothing. Which photos to upload and how many is in getting your actual product into an AI display render.

The format carries its proportions. The display type is chosen from a taxonomy with the proportions attached, so the shelves start at the width and pitch of the format rather than stretching to hold whatever load the composition wanted. The eight types and what each one carries are in what an AI point of sale display generator has to get right.

The brief carries what the client asked for. The briefing step reads the customer's email and brings the format and the must-haves into the brief, so the facings they wrote down travel with the job.

A drawing beats a number. If the client sent a sketch with the shelves drawn in, the sketch route makes the render follow its structure, silhouette and shelf count at the faithful setting. A layout someone drew is a stronger anchor than a figure in a sentence.

The count gets checked on the right picture. On Pro, Generate angles returns a flat elevation, profile and plan of the same display. The elevation is where you count facings across each shelf, and the plan shows how many deep. Remove product, on the free plan, renders the same display empty, so you can count the shelves and judge their width without the load in the way. Compare puts two versions side by side at no cost in credits.

None of that turns a picture into a count. It makes the load your real packs on the right format, and gives you one view where the number reads in seconds, before the client reads it for you. The five-minute version of that check, with the pack as the ruler, is in how accurate are AI display renders.

If the packs for the brief you're working on are in a folder right now, upload them with the brief and count the elevation that comes back.

Where the count comes from

General image modelAI POP DisplaysBellto
The packDrawn from a descriptionYour pack photos, one per SKUThe product, as the input proportions come from
The facingsA word in the promptYour brief and your sketchAn input the piece is designed around
Shelf widthWhatever the composition neededThe proportions of the formatDerived from the product and the facings
How you checkA three quarter viewFlat elevation, profile and plan, plus the empty structurePart-by-part cutlist and dimensioned drawings
What you getA picture of a full shelfA concept with a load you can countA piece built around the count

AI POP Displays runs on Google's Nano Banana Pro. What decides the load is what sits on top of it: the format with its proportions, the real packs as the product, a sketch the render follows and the flat views to check on. That's close to four years in this industry, built into the tool once.

What we're building next: Bellto

A concept render carries the look of the load and a count you can read. When the number has to be exact, because it goes into a quote, a planogram or a retailer's capacity requirement, it stops being a picture problem. An exact count is a shelf built around the pack and the facings, and that's geometry.

That's the half we're building now. Bellto is an AI agent for brands and agencies that takes a POP campaign from brief to shop drawings. You tell it what you're launching (brand, product, channel, stores, budget, date) and it asks for what's missing. It proposes the campaign mix and materials, floor stand, glorifier, shelf strip, stopper, with the budget split per store, and designs every piece with you at real scale in parametric 3D, proportions derived from the product and the facings. The facings are an input the piece is designed around, not something you count afterwards. A dimension change recomputes the whole piece and re-runs the manufacturing checks in seconds. Every approved piece comes out as a 3D model, a part-by-part cutlist, dimensioned drawings, a STEP of the assembly and a cut DXF per part, a package a workshop can quote without redrawing, and it suggests manufacturers that fit by material, format and volume. The quote comes from the manufacturer.

We're building it now. The waitlist is open to any brand, and we're contacting the first ones soon to run the first real campaigns end to end. It opens by invitation, in small groups. Pricing goes first to the people on the list. No card. POP manufacturers can sign up too.

What should you do with the render on your screen?

Count it on a flat view before anyone else does. If the number is wrong, the fix isn't a louder prompt. It's the real packs, the right format and a view you can count on.

For the display on your desk this week, signup is free with 15 one-time credits and no card required, enough to load the real range onto the format and count what comes back. Pro is $49 a month for 150 concept generations at the founding price, with the $69 list price stated openly. Nothing you upload or render trains AI models, on any plan. And if the count has to be exact across a whole campaign, the Bellto waitlist is where it starts.

Frequently asked

Why does AI put the wrong number of products on a display render?

Because an image model doesn't count. It draws a picture that matches the request as a whole, and the number is one weak signal among many. What it learned from retail photography is that shelves look full, so it fills the shelf, and the number of packs comes out of the shelf width and the size the pack happened to be drawn at. A 2025 study titled Text-to-Image Diffusion Models Cannot Count, and Prompt Refinement Cannot Help found that all the state-of-the-art models it tested fail to generate the correct number of objects, with accuracy dropping as the number grows.

Can I fix the product count with a better prompt?

Not reliably. The same study tried rewriting counting prompts in several structured ways and found that these simple interventions generally do not improve counting accuracy. The count improves when it stops depending on a word: the real pack loaded as the product, the format's proportions, a sketch with the shelves drawn in, and a flat view to check the result on. AI POP Displays is built around those inputs.

How do I check the number of products in an AI display render?

Count on a flat front view, never on a three quarter view, because perspective stacks the shelves and hides the back rows. Use the pack as the ruler: count pack widths across each shelf and compare with the shelf width you plan to build divided by the pack width. In AI POP Displays, Generate angles on Pro returns a flat elevation, profile and plan of the same display, and Remove product on the free plan shows the empty structure so you can count shelves without the load in the way.

How do I get a display with the exact number of facings?

From geometry, not from an image. An exact count means the shelf is built around the pack and the facings. Bellto, an AI agent for POP campaigns that we are building, designs every piece at real scale in parametric 3D with proportions derived from the product and the facings, and delivers a part-by-part cutlist, dimensioned drawings, a STEP of the assembly and a cut DXF per part. It opens by invitation from a waitlist.


Start generating Get on the Bellto waitlist