My AI Display Render Came Out Wrong. Now What?

The render came back and something's off. Four shelves where the brief said three. A logo that reads like the brand name spelled by someone who heard it once. Nine packs across a shelf that takes five. A base that would tip the moment a shopper pulled a can from the top tier.

The instinct is to type the same request again, louder. That costs you the parts of the image that were right. A wrong render is a diagnosis before it's a retry.

I've worked in point-of-purchase for close to four years and I build AI POP Displays. The first step is always the same, working out what kind of wrong it is.

What kind of wrong is it?

There are three kinds, and each one gets fixed in a different place.

The brief left it open. The model had to choose a number, a proportion or a material because nobody told it, and it chose badly. This gets fixed in the brief, and then you regenerate.

One thing is wrong on an otherwise right image. The structure is what you wanted, and the background, the graphic, the view or one detail isn't. This gets fixed on the image, with an edit that holds everything else.

The thing that's wrong isn't in the picture at all. A measurement, a load path, a fold. No wording fixes these, because an image model draws what looks plausible and measures nothing. This goes to a drawing.

Search for help with this and you land in one of two places. The first is guides for fixing interior and architectural renders. The second, the moment you type the word shelf, is computer vision that audits photos of real store shelves. Neither is about a generated concept of a fixture that somebody has to quote, which is the one on your screen.

The shelf count doesn't match. What now?

Read your own brief first. If it says "floor display for a soda brand" and nothing about tiers, the model drew the number of shelves that balanced the composition, and it would draw a different number tomorrow.

That one is fixed in the brief. Name the tiers, name the facings across each one, and say what the shelves have in common, as in three equal shelves, eight facings across, two deep. I laid out the full list of what a display brief has to name in how to prompt AI for retail displays, and product load is the line almost everybody leaves out.

If you have a drawing, use it. A customer's sketch is the strongest lock there is on a count, because in AI POP Displays the sketch travels as the last reference and, at the faithful setting, the structure, silhouette and shelf count are locked to it.

If the brief did name the count and the render still came back with four, that's the second kind of wrong, and you don't regenerate. You'd be throwing away a header, a material and a light you already like. On Pro, a custom edit takes one instruction in your own words on the existing image, and the edit is built to hold everything the instruction doesn't name.

Then count on the right picture. A three quarter view is the worst image to count shelves on, because perspective stacks them. On Pro you can generate a flat front elevation of the same display, with no perspective, and count there. If your render came out of a general chat model with a shelf too many, run the same brief through the free plan with the tiers written in and compare the two.

The logo came back warped. Can a better prompt fix it?

No, and this is where people lose the most time trying.

A brand mark isn't text the model has to spell. It's a specific drawing, in a specific color, sitting at an angle across a header card or a curved pack. Describe it in words and the model draws something near it. Reword the description and you get a different something near it. I went through why in why generic image AI fails at POP displays.

What fixes it is the real file, in the right role.

At generation, upload the logo as a reference. In our pipeline every reference carries a written label in front of it, logo, product photo, campaign artwork, style reference or sketch, so the model knows that image is a mark to reproduce and not a mood to borrow from.

On a render that already exists, use apply graphic, which is a Pro edit. You give it the artwork file and it places that artwork on the branded surface where it belongs, the header first, then the front face, side panels and base, in perspective, with the colors of the file kept and the product left uncovered. The display under it is held as it was.

Then look at it at full size, not the thumbnail, because a brand manager catches a soft letterform on sight. The final artwork is still the print file your manufacturer works from. The render's job is to show the client their mark on the unit.

The packs don't fit the shelf. What fixes that?

This is two different failures that look alike.

The first is the wrong pack. The model drew a bottle that resembles your client's bottle, or the same generic carton on every shelf when the range has four SKUs. That's a reference problem. Upload photos of the real packs. They're used as the product, placed at scale, facing forward, with each variant kept separate rather than merged into one.

The second is the wrong number of packs. Too many across, or a large SKU and a small one drawn the same size. That's the brief again, so state the facings, and then check the render with the pack as the ruler. Count pack widths across a shelf and pack heights between two shelves, and compare with what you plan to build. I wrote the whole five-minute check in how accurate are AI display renders.

Two edits help you see what you have. Remove product, which is on the free plan, renders the same display empty, so you can judge the structure without the load hiding it. And a custom edit takes up to four reference images along with the instruction, so "replace the cartons on the top shelf with this pack" arrives with the pack attached.

If the packs on your client's display are on your desk right now, upload them with the brief and count the facings that come back.

The base wouldn't stand up. Can I prompt that away?

You can make it look better. You can't make it stand.

Rewording does move the picture. "Base no narrower than the shelves above it" and "every shelf lands on a side panel" are proportions the model can see and follow, and a render that ignores them usually had nothing in the brief asking for them. In AI POP Displays the format comes from a taxonomy that carries those proportions, so a floor unit starts out taller than it is deep with a base under its load.

What no wording does is calculate anything. The image doesn't know the weight of forty cans or the grade of the board. It shows a base that looks stable, at the same confidence whether it is or not.

So treat this one as the third kind of wrong. Get the look approved, then hand engineering something they can work from. Convert to line drawing gives a clean monochrome drawing of the same display from the same angle, and the flat elevation, profile and plan give them the three views without perspective. Both are Pro edits. Neither is measured. They're the brief for the person who does the measuring.

Should I edit or regenerate?

The rule I use is whether you're changing the answer or the question.

What's wrong	Kind	Where it gets fixed
Shelf or facing count the brief never named	Brief	Write the count in, attach a sketch if you have one, regenerate
Count named in the brief and ignored	Image	Custom edit on the existing render, then count on a flat elevation
Logo or artwork approximate	Image	Logo as a labeled reference, or apply graphic with the real file
Generic or wrong pack	Brief	Upload the real pack photos as the product
Wrong store, wrong crop	Image	Change background or change ratio, both on the free plan
Wrong format or material family	Brief	Change the field and regenerate, it's a new concept
Base, load, folds, millimeters	Neither	Line drawing and flat views to engineering

Two things make editing safe. Every edit lands as a new version next to the one it came from, so the image you liked is never overwritten. And compare puts any two versions side by side at no cost in credits, which is how you check that the structure, the shelves and the branding survived the change. If an edit fails to render, the credit goes back automatically.

The client-facing side of this, where each round of feedback becomes a list of edits, is its own subject and I covered it in how AI changes the client revision round.

What never gets fixed by prompting?

Anything with a number behind it. Millimeters, because nothing in the image has a length. Whether the unit stands, folds and ships, because a render shows a fold line where one would plausibly be and doesn't know if the tab locks. A retailer's footprint, because a bay or a pallet position is a dimension, and dimensions go in the document your manufacturer quotes from.

AI POP Displays runs on Google's Nano Banana Pro. What keeps a wrong render quick to fix is what sits on top: a display taxonomy with proportions, materials that behave, retail environments as options, product load from the real packs, references labeled by role with the sketch last, a briefing step that reads the customer's email, and an edit console that holds the display while one thing changes. That's close to four years in point-of-purchase, built into the tool once.

Where the third kind of fix is going

The failures in that last section all have the same cause. They need geometry, and a picture has none. 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. 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, and the campaign you describe when you sign up sets your place. Pricing goes first to the people on the list. No card. POP manufacturers can sign up too.

For the render that came out wrong this morning, signup is free with 15 one-time credits and no card required, enough to put the brief back through with the count, the packs and the logo in it. 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.