AI-Generated Marketing Photos: A Cafe Test of What Looks Honest

They are safe when the picture still shows what a customer will actually be served, and risky when it does not. Light AI edits to a real photo, such as exposure, crop or removing a bin from the background, are fine. A fully generated dish is not honest for a menu, a delivery app or a Google profile.

One menu item in three versions makes the line visible: a real photo, the same photo edited with AI, and a fully generated image. What decides honesty is the plate, not the tool. If a customer could hold the photo next to what arrives and feel misled, the photo fails, however it was made. Platforms add their own rules on top, and Google's photo tips for business profiles now name AI outright.

Follow me on Instagram@sagnikteaches

The test: one dish, three versions, ten customers

The cafe in this illustration has 28 seats and a short brunch menu. Its best-seller is smashed avocado on sourdough with a poached egg and chilli flakes, at $13.50. The owner wants one photo of it for the website, the delivery app and Instagram, and makes it three ways.

Connect on LinkedInSagnik Bhattacharya

As of October 2026, list prices in this area include Photoroom Pro at $12.99 a month ($7.50 a month billed yearly, with 1,000 exports) and Adobe Firefly Standard at $9.99 a month (2,000 credits). Canva's free plan gives a small monthly allowance of AI uses, up to 20 at the last check. The Gemini app's free image downloads are 1K size with a visible sparkle watermark, and 2K comes with an AI plan. Plans change, so check before you rely on one.

Subscribe on YouTube@sagnikteaches

Three questions decide the result: how closely the picture matches the plate, whether a customer would feel misled, and whether the platform would accept it.

Version A: a real photo, shot properly

A real photo is the only version that cannot be called misleading by accident, though it can still mislead by styling: a hero plate with extra avocado is no more honest than a generated one. A few habits keep it fair.

• Light. Side light from a window, with the warm lamp off. Overhead bulbs flatten the food and cast a yellow tint.
• Angle. About 45 degrees, which is how a seated customer sees the plate. Straight overhead suits flat dishes but hides height.
• The plate as served. Straight from the pass, on the crockery you use, with the rim wiped. No extra garnish for the camera.
• Twelve frames, then choose the typical one. Pick the frame whose portion matches an average plate, not the most flattering plate of the session.
• A size cue. A cup or fork in the frame shows scale.
• Frame wide, in landscape. DoorDash's rejection guide accepts only 16:9 landscape and warns against cropping a vertical photo to fit, while an Instagram portrait post is 3:4. Leave space around the plate and one shot can be cropped both ways without cutting into the food.
• The delivered dish too. If the item travels in a box, photograph it as it is packed. Delivery customers judge the box, not the plate.

Version B: an edit ledger for what AI may change

Version B starts from the real photo, so the question is which edits keep it honest. Write the answer down as a ledger and apply it every time, whoever edits.

Watch the tools that regenerate part of a picture. Photoroom's AI Product Fixer, for example, regenerates whichever area you brush over and leaves the rest intact. That may suit a product shot, but it is risky for a plate of food, because it can quietly change what is on the plate. After any edit, open the original and the edited version side by side at 100% zoom and count the eggs, the slices and the flakes. The comparison takes a minute and catches the one change that matters.

Here is a mistake that one-click retouch buttons invite. The owner taps a "fix" button on the egg, which looked slightly flat. The tool redraws the yolk. In the edited image the yolk is rounder, glossier and larger than the one in the kitchen, and nobody notices until the side-by-side check. The ledger would have refused the edit under "make the yolk runnier or more golden". The remedy is to go back to the original and use exposure alone.

The ledger is written as edits, not as dishes, so it carries over to the next item without a new debate. Take the pastry cabinet. A carrot cake slice photographed through glass picks up reflections of the window and the queue. Removing the reflections, straightening the shot and cropping all pass. A thicker layer of frosting, or a whole cake added behind the slice, does not.

Version C: what a generated plate gets wrong

A prompt like the following is where most people start.

Overhead photo of smashed avocado on sourdough with a poached egg and chilli
flakes on a white plate, natural window light, wooden table, busy cafe background

The faults to look for are predictable. The model draws an average avocado toast, not yours, so expect extras: a second egg, a fan of sliced avocado instead of a smashed spread, microgreens or edible flowers the kitchen does not use, a perfect glossy yolk and a plate you do not own. Asking again with "one egg, smashed avocado, no garnish" fixes some of it and breaks something else, such as toast that turns thin and white. There is no way to edit it back into your plate, because it never was your plate. Each retry draws another plausible one.

Free-tier output can also be small and carry a visible watermark, which tells customers, and a delivery app's reviewers, what they are looking at. Sizes and marks vary by tool and plan, so check the one you use. The honest uses of version C are a clearly labelled illustration and atmosphere graphics, such as a seasonal poster, where nobody is being told "this is what you will get".

What customers expect from a menu photo

Ask what a customer expects and the answers are modest. They expect the same components on the plate, in roughly the same quantity, on roughly the same crockery, looking like the dish on a good day. They accept flattering light, a tidy rim and a sensible crop. What they resent is the gap: an extra egg, a missing side, a portion that looks twice the size, a garnish the kitchen has never heard of. Delivery customers add a second expectation, that the picture shows the dish in the form that arrives, so a plated hero shot of something that comes in a box sets up a disappointment before the lid is off. That is the standard the customer test turns into numbers, and it is close to the one the platforms apply when they say a photo must represent the item.

The customer test: would they expect this plate?

Ask ten people, regulars and staff who do not make the dish, to look at the three images on a tablet at the same size, in random order, with the real plate on the table in front of them. A quiet weekday morning works, and it takes about ten minutes a person; buy them a coffee as thanks, and keep that separate from any request to leave a review. Three questions each. How close is the picture to this plate, from 1 to 5? If you had ordered from this picture, would you feel misled, yes or no? And which of the three do you think is the real photo?

The figures below are invented to show how to read the sheet. They are not survey results.

My suggested pass mark is a mean of at least 4.0 and no more than one person in ten feeling misled. On those numbers C fails by a distance, and it fails on the question that matters even though only one person took it for a real photo. B passes, but the one person who felt misled said the colours looked brighter than the plate, so the owner lowers the saturation and re-runs the check with a different ten. The third question is the quiet one: if many people pick the generated image as the real photo, it is convincing, which is exactly why it needs a label.

Customers will also post their own photos of the plate on Google and Instagram, and those will sit beside yours. A menu photo that does not match invites the comparison. The guide to fair and misleading edits for hotel photos applies the same test to a different business.

Platform rules that apply to a cafe's photos

Rules differ by platform and move quickly, so treat this table as a list of questions to check on each platform's current page, with what the pages said in October 2026.

The pattern across platforms is consistent: edits that leave the food as it is are accepted, and images that no longer represent the item are not. The wider routine for restaurants is in using AI food photography on menus and delivery apps, and the labelling side is in labelling AI-generated images and video on social media.

Advertising rules in plain terms

This is general information, not legal advice, and I have left out named regulators because the right one depends on where you trade. Four points hold in most markets.

• Overall impression. Advertising rules in many markets judge an image by the impression it gives a customer, not by the small print. A photo that suggests a bigger portion, more toppings or a finer finish than the customer gets can count as misleading, whoever or whatever produced it.
• Captions need backing. "Homemade", "freshly baked", "locally sourced", "organic", "vegan" and "free-range" must be true of the dish in the picture. An assistant that writes your caption will happily upgrade a claim, so compare every adjective with your recipe and supplier notes.
• People need consent. Real staff and customers in a photo need to agree to it. A generated "happy customer" or "our barista" is not a customer or a barista, and presenting one as real is a misleading testimonial.
• Labels. If you have customers in the EU, the AI Act's transparency rules for AI-generated and manipulated content have applied since 2 August 2026. Whether a generated image of a dish falls inside them is a question for an adviser, and a label costs you nothing, so label it.

The longer legal picture for marketing images is in whether you can legally use AI-generated images in your marketing.

Captions and records that keep you honest

Three lines of caption wording cover most posts.

Real photo, no edits beyond crop and light: no label needed.
Edited real photo (background tidied): optional, "Photo edited for light and crop."
Generated image or illustration: "Illustration made with AI. Our real plate is
smashed avocado on sourdough with one poached egg and chilli flakes."

Keep a record for every image you publish, so that a platform query or a customer complaint takes two minutes to answer. One row per image is enough.

For generated images, add the prompt and the tool's settings. For the quality check before posting any of them, checking AI-generated images for errors before posting lists the faults that slip through, such as garbled text on a sign or a hand with the wrong number of fingers.

Three weeks from first shot to published photo

1. Week 1. Shoot version A and choose the frame. Make version B using the ledger, and keep the original. Decision: does B pass the side-by-side check?
2. Week 2. Make version C only if you want a labelled illustration. Run the customer test with ten people and score the three images. Decision: do A and B reach the pass mark?
3. Week 3. Check each platform's current photo page. Upload the winning version with the right caption, store the record row, and check how it displays on a phone. Decision: are you happy to hold the photo up next to a real plate in front of a customer?

Which version for which job

For this cafe the outcome is simple. Version B goes on the menu pages and delivery apps, version A is used for Instagram close-ups, and generated images appear only on a seasonal graphic marked as an illustration. The test took about three weeks of evenings, cost at most one month of an editing app plus ten coffees, and left a ledger, a record sheet and a pass mark the next photo can use. The same routine works for any menu item, and the coffee shop social media routine shows where the real photos fit into a month of posts.

Questions cafe owners ask about AI food photos

Can a cafe use AI-generated photos of its own dishes?

Not for the dishes themselves. A generated image shows an average plate, not yours. DoorDash, for one, rejects photos that show an idealised or generic version instead of the item served, and Google's photo tips warn against excessive use of AI. Use real photos, lightly edited, and keep generated images for clearly labelled illustrations or atmosphere graphics.

Is it fine to use AI to tidy the background of a real food photo?

Usually, if the real background stays and the food is untouched: removing a bin, fixing light and cropping are ordinary edits. Replacing the background with a different scene is riskier, since delivery apps list creative backgrounds among rejection reasons. Keep the original file and a note of every edit.

Do I have to label AI-edited photos?

Platform rules differ and change, so check each one's current help page. Labelling a fully generated image costs nothing and is the safe habit; a light edit for exposure or crop is not normally labelled. If you have customers in the EU, ask an adviser whether the AI Act's transparency rules reach your images.

What is the quickest honesty check for a food photo?

Put the photo beside the real plate and ask someone who has not seen the dish whether they would feel misled if the plate arrived looking like that. If the answer is yes for any reason, such as portion, toppings, colour or crockery, retake it or edit less.

Further reads

• AI Product Photos vs a Photographer: Cost and Quality Compared — When a photographer still beats AI for product shots.
• How to Create Product Photos With AI on a Small Budget — Making clean product photos on a small budget.
• Brand Guidelines for AI Images: Colours, Style and Limits — Setting colour, style and limits for any AI images you use.
• Takeaway Delivery App Menus: Rewrite Descriptions and Photos With AI — Delivery app menus: descriptions and photos together.
• AI for Used Car Dealerships: Honest Listings and Photo Clean-Up — Another trade where photo clean-up has to stay honest.
• ChatGPT Prompts for Coffee Shop Owners: 10 for Menus, Posts, Rotas — Ten ChatGPT prompts for a coffee shop, from cost per cup to rotas and posts, each with example inputs and a check before you use it.
• Starting a Business Using AI: A First-Month Plan for a Shop or Cafe — Four weeks before opening: names, supplier emails, menu or range, a cash-flow draft and launch posts, with the checks a person must do.
• AI Marketing Campaign Examples: Five Worked Ones for Local Businesses — Five worked local campaigns, each with a prompt, assets, schedule, cost, break-even point and a simple way to check whether it paid for itself.
• AI for Takeaways: What Is Worth Automating and What Is Not — Four takeaway jobs ranked by effort and payback, and the ones to keep human: allergy answers, refunds, pricing and kitchen capacity.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: DoorDash's merchant help pages on rejected menu photos and its smart photo tools; Google's Business Profile photo tips; Google's merchant image metadata rules; Meta and TikTok help pages on AI labels; Photoroom, Adobe Firefly and Canva pricing pages and Google's Gemini image help, as checked in October 2026.