Physix Frontier
Why Menu Burgers Look Less Like Real Burgers These Days
Last week, a friend who runs a small restaurant in St Louis sent me a WeChat message saying delivery platforms were urging him to add dish photos. His shop had no budget for photography, the menu was all text, and the default gray images looked like the place wasn't operating. He asked, since AI image generation is so convenient now, could he just generate a few burgers and fries to fill the gap?
This pain point is very common. Small shops need to survive, and first they need images. But once the images are fake, before the food is even served, trust collapses halfway. I've recently been using large models for image generation, for about three weeks, so I decided to take his most ordinary burger and fries order and run it through to see if these so-called "easy menu photos" are actually usable.
First, I went to the shop and took a set of photos. At 3 PM, the light by the window was hardest. The burger came out slightly collapsed, and the oil sheen on the fries had settled. After shooting with my phone, I simply cropped and adjusted brightness, taking about forty minutes. The photos weren't beautiful, but there was an obvious benefit: I knew they were edible today.
Back home, I tried the second approach: generating menu images with an image model. The interface was standard: prompt on the left, aspect ratio and style selection on the right, click generate and wait ten seconds. Initially, I wrote it plainly: "St Louis diner burger, golden fries, warm light, realistic menu photo." The first result came out, and I stared at it for a long time. It was bright, yes, but it didn't look like food.
The sesame seeds on the bun looked stuck on, every fry was identical in thickness, the ketchup reflection was too uniform, like polished plastic. The background was a non-existent little eatery, with text on the wall that blurred into symbols when zoomed in. The sticking point was obvious: the model is good at "drawing menus" but doesn't quite understand "menu photography." It always pulls the image toward high-end ad styles: black backgrounds, strong outlines, oversaturation. I repeatedly tweaked prompts, adding "no excessive retouching," "realistic oil sheen," "napkins on table," "slight steam," and only got one barely acceptable image on the third round.
There were surprises too. Generation speed was fast; a set could be produced in ten minutes. My friend's shop had one empty table with a particularly messy background. I used the model to swap the tabletop for wood grain and move the cups away, making the whole scene much cleaner. For temporary image filling, this efficiency is indeed tempting.
But I didn't dare use the AI images directly for him. I did a small test, picking one real photo and one generated photo each, showing them to six people in his shop. Not many people, so it's just a clue, not data. I asked them which one made them more willing to order. Five chose the real photo, reasons similar: it looked "edible today," while the AI one looked "ad-heavy." The remaining one chose AI because the image was cleaner, without grease stains or wrinkles.
The second question was more direct: if you knew this was AI-generated, would you feel misled? Most said they wouldn't be angry, but expectations would drop, and if the actual dish differed too much, they'd still leave a bad review. This reaction aligns with recent reports. St Louis restaurants using AI for food images faced diner backlash. Media called such images "slop," meaning blurry and fake, not looking edible. The Guardian and NBC reported similarly: consumers find AI menu images unappealing, even misleading.
I checked several reports; it's not just single-platform venting. Samples varied across reports, involving LA night markets, SF coffee shops, and specifically St Louis local restaurants. It shows resentment has become a universal intuition. Diners resist "fake images presented as real promises."
My final conclusion is: it depends. It can be used for filling blank slots, keeping delivery platforms from looking empty; if owners lack budget for photography, it can serve as a transitional solution, provided "images are for reference only" is explicitly stated. It's unsuitable for community shops relying on regulars, places where dish presentation varies, or shops already questioned by customers for "image mismatch."
Pros are straightforward: fast, fills gaps, swaps backgrounds, cleans up messy tables. Cons are hard: details look fake, highlights look weird, food doesn't look like food, restaurants don't look like restaurants. Worse, it amplifies existing customer suspicion. Your burger looks great in the photo, arrives collapsed, and customers won't blame the model—they'll blame you.
Menu photos are a small contract.
My friend didn't use AI entirely. He picked one phone-shot real photo as the main image, demoted an AI image to background lighting filler, and added a note on the platform: "Images taken on-site, slight lighting differences." Regulars entering the shop still ask first: "Is today's burger like the one in the picture?" He said at least he can answer that honestly.
📌 This article is compiled from Hacker News, original address: https://www.stltoday.com/news/local/business/article_4049a9c9-e801-4073-a8a0-347494790282.html
Copyright belongs to the original authors. This is a compilation and independent analysis based on public reports.