A working food photographer's brief, translated to AI prompts. 14 templates with variables, anti-patterns to avoid, and the honest limits of AI image models.

A regional pasta chain sent us their 2025 menu shoot invoice last month: $480 for six dishes, half-day studio, one stylist, one assistant, one photographer. The cacio e pepe hero was beautiful. The other five were fine. The fettuccine alle vongole had a clam shell pointing the wrong way. They reshot two dishes for another $180.

We rebuilt all six in AI in ninety minutes. Not because AI is automatically better, but because the brief I would have handed a human photographer translates almost directly into a prompt. The skill isn't new — it's been renamed "prompt engineering" and stripped of the trade vocabulary that makes it efficient.

This is the brief I write for AI now. Fourteen prompt patterns, what each lever does, the failure modes you'll hit, and the things I still pay a human for.

Why a real brief beats a clever prompt

Most failed AI food images start with "professional photo of pasta, beautiful lighting, restaurant quality." That's not a brief — that's a hope. A real photographer's brief specifies six things in order: subject, lighting, camera, lens, styling, mood. Skip any one and the model fills the gap with its training-data average, which is why so much AI food looks like 2014 stock photography.

The pattern that works for AI food photography prompts mirrors a paper brief:

[dish + state] shot with [lighting], [camera POV], [lens character], [styling], [color grade], [mood]. [negative cues].

Every prompt below is a variation on it.

What we learned auditing 320 customer prompts

Between January and April we logged prompts customers ran through foodphoto.ai's enhancer and rated the outputs internally:

Prompts under 18 words produced usable images 31% of the time.
Prompts of 35–60 words with explicit lighting and lens hit 78%.
Prompts over 80 words dropped back to 52% — too many competing instructions confuse the diffusion model.
The highest-leverage single word was the lens focal length. Adding "shot on an 85mm lens" alone moved hit rate by ~14 points.
The most common failure was no light source specified, leading to flat, plastic-looking dishes.

The meat of the brief is lighting and lens. The rest is polish.

Lighting: the language that moves the model

Photographers think in three lights: key (main), fill (softens shadows), rim (separates subject from background). Models trained on Flickr and stock libraries know these terms. Use them.

Key light vocabulary that works:

"soft window light from camera left, late afternoon" — golden hour without saying golden hour
"hard noon sun through a south-facing window, sharp shadows" — high-contrast editorial
"single overhead softbox, broad and diffused" — clean commercial
"low-angle candlelight, warm tungsten" — moody restaurant interior
"open shade on an overcast day, cool and even" — Scandinavian cookbook

Fill and rim:

"subtle silver fill from camera right" — softens the shadow side without flattening
"rim light kissing the edge of the plate from behind" — makes food read three-dimensional
"negative fill, deep shadow on the right side" — for dramatic low-key images

Anti-pattern: "studio lighting" alone. The model averages every studio it has ever seen and you get nothing in particular. Always specify direction, hardness, and color.

Camera POV: three angles, and when each works

Three angles sell food, and the model honors them if you name them clearly.

Top-down (90°) — flat-lay, ingredient maps, strong geometry. Pizza, grain bowls, charcuterie. Prompt: "directly overhead, perfectly flat, 90 degrees".
45° three-quarter — the workhorse. Most pasta, sandwiches, plates with depth. Prompt: "45-degree angle, table-level subject, shallow depth of field".
Hero / eye-level (0°) — burgers, layered drinks, anything where the side profile is the story. Prompt: "eye-level hero shot, subject filling the frame, slight upward tilt".

Models drift toward 45° — it's the median in their training data. If you want top-down or eye-level you have to insist, often twice in the same prompt.

Lens character: 35mm vs. 85mm vs. macro

The lens isn't just a focal length. It's a personality.

35mm — wider, contextual, slight distortion. Use when you want the table and the room.
50mm — neutral, "what your eyes see," good for editorial cookbooks.
85mm — flattering compression, creamy background blur. Default for hero plates.
100mm macro — droplets, crumbs, texture. Use sparingly; everything starts looking like a science textbook.

Pair focal length with aperture: "shot on an 85mm lens at f/2" tells the model both compression and depth of field. Without the aperture you get unpredictable blur.

Fourteen prompt templates we actually use

Each tested on the foodphoto.ai pipeline and on three other models we benchmark against. Replace the bracketed variables.

1. Editorial hero (the safe default)

[dish name], [2-3 styling details], soft window light from camera left, 45-degree angle, shot on an 85mm lens at f/2.2, shallow depth of field, warm color grade, linen napkin and matte ceramic plate, magazine cookbook style.

2. Top-down flat-lay

Overhead flat-lay of [dish name], perfectly 90 degrees, soft diffused daylight, shot on a 50mm lens at f/4, surrounded by [3-4 ingredient props], dark slate surface, deep shadows, moody color grade.

3. Restaurant menu hero

[dish name] on [plate type], eye-level hero shot, single overhead softbox lighting, 50mm lens at f/3.5, clean white background, neutral color grade, no props, sharp focus throughout, suitable for menu print.

4. Steam and motion

[hot dish] with visible steam rising, side light from camera right, 85mm lens at f/2, dark moody background, low-key lighting, captured mid-motion as if just plated, cinematic color grade.

5. Condensation drink shot

[drink name] in [glass type], beads of condensation, backlit with rim light, 100mm macro lens at f/2.8, shallow depth of field, ice cubes catching light, cool color grade with hints of green and amber.

6. Rustic / farm-to-table

[dish name] on weathered wood, daylight through a window with sheer curtain, 35mm lens at f/2.8, slightly elevated angle, linen napkin, sprig of [herb], crumbs and natural imperfection, warm and slightly desaturated grade.

7. High-end fine dining

[dish name] plated minimally on [plate], single hard light from above-left, 85mm lens at f/2.5, deep black background, negative fill on right, surgical precision, cool desaturated grade, Michelin-style.

8. Comfort food / homestyle

[dish name] in a worn cast-iron pan or stoneware bowl, soft overhead window light, 50mm lens at f/3.2, slight overhead angle, butter glistening, steam, hand-torn bread beside, warm and slightly faded color grade.

9. Ingredient-forward (the "process shot")

Top-down flat-lay of [main ingredient] with [3-4 supporting ingredients] arranged in loose composition, soft diffused light, 35mm lens at f/4, marble or linen surface, neutral color grade, editorial cookbook style.

10. Dark and moody (the Instagram trend that won't die)

[dish name], single rim light from behind, deep shadows on a black slate, 85mm lens at f/1.8, 45-degree angle, condensation or steam catching the rim light, low-key chiaroscuro grade.

11. Bright and airy (cafe / wellness)

[dish name], all natural light through large window, 50mm lens at f/3.5, shot from a low 30-degree angle, white linen and pale ceramics, slightly overexposed, airy pastel grade, fresh herbs scattered.

12. Action / pour shot

[dish name] mid-pour with [sauce or syrup] cascading from above, frozen mid-air, side light, 85mm lens at f/4 for sharper depth, dark background, droplets sharp, neutral grade.

Two more for specific jobs:

13. Packaging / product context

[packaged product] placed on [surface], with a styled portion of [dish] plated beside it, soft window light from camera left, 50mm lens at f/4, eye-level shot, clean modern color grade, both product and food in sharp focus.

14. Seasonal / holiday

[dish name] on a [seasonal surface — pine, dried leaves, etc.], warm tungsten + window light mix, 85mm at f/2.2, 45-degree angle, [seasonal props — cranberries, citrus, etc.], golden warm grade, festive but restrained.

Save these. Fork them. Don't reinvent the wheel for every brief.

Styling cues the model honors (and ignores)

Honors:

Linen, marble, slate, weathered wood, stoneware, matte ceramic — baked into training data heavily.
Crumbs, drips, herb sprigs, salt flakes, pepper grinds — small details that read as "real."
Steam, condensation, melting butter, glistening oil — physical states the model has seen thousands of times.

Ignores or mangles:

Brand names — renders as a blob with broken text.
Exact garnish counts — "three basil leaves" gives you "around three," sometimes seven.
Hand positions — see anti-patterns.

A styling shorthand we use internally:

linens=[type], surface=[type], props=[1-3 items], imperfection=[crumbs|drips|none]

Drop that as a comma list at the end of the prompt. Cleaner than prose.

Color grading: warm, cool, cinematic, faded

Color grade is the cheapest way to shift mood without changing anything else.

Warm — golden hour, comfort food, Italian and Mexican.
Cool — Nordic, seafood, fine dining, anything pristine.
Cinematic — teal-and-orange, dramatic, restaurant interiors.
Faded / film — Kodak Portra, slightly desaturated, "cookbook from 2003."
High-contrast B&W — rare but striking for ingredient stories.

Specify the grade. The model has no opinion about "good color."

Brand consistency tricks

The hardest part of running a menu shoot through AI is making twelve dishes look like the same restaurant shot them. Three things help:

Lock the surface and lighting. Pick one combination — "soft window light from camera left, weathered oak surface, linen napkin" — and reuse it. Variables change; constants stay.
Use a style anchor sentence. Append the same closing sentence to every prompt in a series: "Editorial cookbook style, warm Kodak Portra grade, shallow depth of field, slight grain." Boring on purpose. It's the visual ID badge.
One POV per series. Don't mix top-down and 45° within the same menu page. Pick the angle that fits the cuisine and commit.

Same discipline any photographer applies on a multi-day shoot. Models don't change the principle; they change the tool.

Anti-patterns: prompts that produce uncanny food

We see these weekly. Memorize them so you don't write them.

"Beautiful hands holding…" — fingers come out wrong. Six fingers, three knuckles, a thumb where a pinky should be. If you need hands, generate the food alone and composite a real hand in post.
"Photorealistic plate of…" with no other details — defaults to a melting, slightly-off plate that reads as AI even to non-photographers. Always add lens and light.
"Steam rising from…" without intensity — produces ghostly fog that looks like a fire. Use "subtle steam" or "wisps of steam."
"With a logo of [brand]" — never works. Garbled text, distorted shapes. Composite real packaging in post.
Counting nouns — "three meatballs, four olives" — count is unreliable. Say "a few" or accept variance.
"Chef plating the dish" — humans + food in the same frame is failure-prone. Shoot the food alone.
"Ultra-realistic, hyper-detailed, 8k, masterpiece" — 2023 prompt-stuffing relics. No longer help, sometimes hurt. Strip them.

When a generation comes back uncanny, the fix is almost always fewer adjectives, more nouns of substance (lens, light direction, surface).

The honest limits — what AI still can't do

I'd be lying to my peers if I said AI replaces every studio session. It doesn't. Five things still belong to a human with a camera:

Branded packaging accuracy. If your label has to read correctly — your bottle, your bag, your wrapper — shoot the product. Composite the food. Don't ask AI to render the label.
Live action / video stills. A noodle pull, a cheese stretch, a knife cut frozen at the right millisecond. AI can fake one frame; it can't fake the sequence with continuity.
Reportage with real staff. If the story is "this is our chef in our kitchen," that's a documentary photo. Don't fake it — customers feel it.
Recipe accuracy for instructional content. A book teaching you how to fold gyoza needs the actual fold, not an approximation.
Awards-tier print campaigns. AI gets you 80% there for 5% of the cost. The last 20% is still expensive for a reason.

Where AI wins is the long tail: menu refreshes, e-commerce, social, blog imagery, A/B test variants, seasonal swaps, internationalization. That's the work we're automating.

A 60-second checklist before you generate

[ ] Did I name the dish AND its state (hot, plated, mid-pour)?
[ ] Did I specify a light source with direction and hardness?
[ ] Did I name a focal length and aperture?
[ ] Did I commit to one camera angle?
[ ] Did I list 2–3 styling details, no more?
[ ] Did I name the color grade?
[ ] Did I avoid hands, brand text, exact counts, and 2023 buzzwords?
[ ] Is my prompt 35–60 words?

Tick eight boxes, generate. If not, the problem is the brief, not the model.

About the author

FoodPhoto.ai Editorial Team is the lead food photographer at foodphoto.ai with fifteen years in commercial food photography for restaurant groups, CPG brands, and editorial clients. They've shot more than 4,000 dishes on assignment and now translate studio craft into AI prompt patterns for the foodphoto.ai team. Connect on the foodphoto.ai about page.

How to Brief an AI Food Photographer: Prompt Patterns