A cafe owner in Portland asked me to help with her menu board photos last month. She had forty dishes shot on a phone propped against a sugar dispenser, with the afternoon sun blasting through the front window. Half the shots were beautiful. The other half looked like they were taken inside a cave.
What she needed wasn't a better prompt. It was a consistent recipe for lighting and framing that she could describe the same way every time, so the board didn't look like it was assembled from three different restaurants. That's the piece most prompt guides skip. If you want the full picture of what these tools can do first, the AI image generator overview is a decent starting point, but come back — the specifics below are what actually keep a set cohesive.
Write the light before you write the food
Most people lead with the subject. "A bowl of ramen, steam rising, chopsticks." That sentence has no light source in it, so the model invents one, and it will invent a different one next time you generate. Lead with light instead: "soft window light from the left, late afternoon, warm fill on the shadow side."
Once light is fixed, everything else in the image stabilizes. Shadows fall the same way across the whole set. Highlight positions match. When you assemble the final menu board, the photos read as a series rather than a collage.
A practical format I use for every food shot now:
[light direction and quality] + [surface] + [subject] + [camera angle] + [small imperfection]
So: "soft window light from the left, late afternoon, warm fill, matte ceramic bowl on dark walnut, ramen with soft egg and scallion, low three-quarter angle, a few drops of broth on the rim." The imperfection matters more than you'd think. Perfectly clean food reads as fake.
Pin the camera to one or two positions
Menu boards die from angle drift. One dish shot straight down, the next at eye level, the third tilted. Pick two angles maximum for a set and name them explicitly in every prompt: "flat overhead" and "low three-quarter, roughly 20 degrees above the surface."
I'd skip focal length language like "50mm f/1.8" unless you're generating something photorealistic on purpose. For most casual restaurant work, "slight background blur" communicates the same intent with fewer weird artifacts.
One caveat: if you're generating dishes that don't exist yet, keep the plate and table constant across the batch. Same "matte off-white bowl," same "dark walnut surface." That repetition is what your eye reads as a brand.
Handling sauces, steam, and other trouble spots
Sauces blob. Steam turns into cotton. Cheese melts into a plastic sheet. These three account for most of the "looks AI" reaction I hear from restaurant people.
Specifics that help:
Instead of "steam rising," try "faint wisps of steam catching the light, thin and wispy." Instead of "cheese pull," try "melted cheese with visible stretch, uneven edges." Instead of "sauce drizzle," try "thin sauce drizzle with a few uneven drips on the rim."
The rule: describe the visual symptom, not the effect. Models can render a wispy curl of white. They struggle with the abstract idea of "steam."
What to generate and what to shoot
Here's the honest split I recommend. Real dishes the restaurant actually serves? Photograph them, then use AI for cleanup, relighting, or background replacement. Dishes not yet on the menu, promotional concepts, or seasonal specials you're testing? Generate them fully.
Mixing the two is fine as long as light direction and angle match. A generated concept shot next to a real photo looks seamless if the shadows fall the same way. It looks like a mistake if they don't.
One more thing worth mentioning: menu boards have a lot of text on them. Never let the model render your prices or dish names. Generate the image with clean space where the text will go, then set type in your layout tool. AI lettering on a menu is the fastest way to make a restaurant look careless.
A quick workflow that holds up
Start with one hero dish. Iterate until the light, angle, and surface all read the way you want. Save that exact prompt text as a template, and only swap the subject and the small imperfection clause for every following dish. Generate two or three variations per dish and pick one — never generate one and move on.
Then assemble everything at actual menu size before you judge it. Photos that look great at full screen can turn to mush at 3 inches on a laminated card. Check contrast against the paper color too, because dark food on dark backgrounds disappears on cheap printing.
If you want to start experimenting with these recipes on your own set, the image generation tool on the main page handles the prompt-to-image step fine. The rest is repetition and a fixed template, which is the least glamorous and most effective advice I can give. The cafe owner's board went out two weeks ago and she says three people asked about the ramen specifically. Small win, but that's the whole point.