Complete Meal

In these 30 photographs, I trained a neural network—and asked it to present things that never appear in a cafeteria—such as oysters, ramen, wedding cakes, and New Year’s holiday tables—in a cafeteria style.

The food in the frame is determined not only by what is on the plate but also by its presentation: the dishware, the background, the angle, and the lighting. The cafeteria tray is one of the most recognizable presentations—anyone who has attended a school or university recognizes it at a glance.I was interested in testing whether an AI could be taught presentation style rather than specific dishes. If the model learns this presentation style, it will be able to transform anything—from a burger to oysters—into a complete meal.

The training set included 30 food photographs—specifically, only cafeteria lunches. They all share common characteristics:
• Strict overhead shooting angle;
• Bright plastic tray—orange, yellow, green, red, or burgundy;
• White plate, tea or compote mug, fork, and spoon;
• Familiar assortment: soup, main course, salad, bread, and sometimes a banana or apple;
• Even light from ceiling lamps.

How the selection process was prepared: The photos were converted from HEIC (since they were taken simply on my phone) to JPG and cropped into a square. For each image, the BLIP model generated a caption, and the style trigger—"top-down photo of a canteen lunch tray in KOMPLEX style"—was added to the beginning of the caption.

The model performed best with dishes easily placed on cafeteria trays: it retained the original dish but placed it on a colorful tray, added tea, white plates, and cutlery. In all the prompts, the first part was the same—the style trigger.

Interestingly, the model often places more plates on the tray than were in the dataset—8–10 instead of the usual 3–4. It learned not a precise number, but the feeling of the tray being full.

Issues arose with the oysters and the wedding cake. The first attempts were weak because the table setting style was missing. I rewrote the prompts to explicitly list the table setting details, and I added «wooden table, many plates» to the negative prompt.

Not everything worked out, and my experiment was unsuccessful in some areas. The rule I derived for achieving more «table-ready» presentations is: style wins when a dish can be plated across white plates. If a dish has its own strong presentation form (like a dish with ice or a tube), the model chooses only one.

For each primary series prompt, I generated two versions using the same seed: one in base SDXL (the prompt without the words «in KOMPLEX style») and one in SDXL with a trained LoRA. The difference illustrates exactly what the training added.