Spec Work · September 2026
The $0.20 Food Photo
Twenty-four Korean fusion dishes nobody cooked, shot by twenty-four people who do not exist. Every frame is the same photo: food held up to the lens, your own shoes on the ground below. It is the most-posted composition in food social, and the one AI stills get right — no face to hold together, no motion to break.
The formula
Every photo below is this block, plus four lines of casting:
Raw unedited vertical 9:16 iPhone photo, POV looking straight down: one hand holds the food up close to the lens, their own lap, outfit and shoes on the ground below. Flat native iPhone HDR color, natural light, no beauty filter, no film grain, no color grade, no cinematic bokeh. Five fingers, correct anatomy. No text, no logos. Food: [cross-section, ingredient by ingredient, with its wrapper] Hand: [skin tone, nails, one piece of jewelry] Location: [city, the exact ground surface, one piece of litter or furniture] Outfit: [trousers, shoes, one detail]
No reference images anywhere. Describing the hand, the shoes and the floor is enough to cast a different person in every frame.
Twelve kimbap nobody sells
Different hand, different city, different shoes each time.
Twelve jjimdak, all in Korea
A braised dish you cannot hold, so “hold it to the lens” became “lift it on chopsticks”. Every setting is a Korean restaurant or home.
What I learned
| Cast by hand, not by face | Skin tone, nails, one ring and one pair of shoes recast every frame with no reference image. These models block realistic face references anyway. |
| Negatives beat adjectives | “No beauty filter, no color grade, no cinematic bokeh” did more for realism than describing the light. Ask for “cinematic” and you get a render. |
| Name the floor | The ground sells the city better than the skyline. Wet subway grate, talavera tile, beach sand, diner checkerboard. |
| Text still fails | Two frames carry invented Korean label text on soju bottles. Fine at thumbnail, wrong under zoom. |
The receipt
| Item | Cost |
|---|---|
| 2 video clips — Seedance 2.5 | $2.40 |
| 24 stills — Gemini Image 3 Pro | $4.80 |
| Total, 26 assets | $7.20 |
Stills are 20 credits each at 1536×2752, four at a time, about ninety seconds a batch. One credit is one cent, billed through the Runway API. All 24 stills came back usable on the first attempt.
The verdict
A twelve-post food series that would need a stylist, a location day and a hand model comes back in four minutes for the price of a coffee. What still needs hands: picking which dishes are funny, writing the casting lines so the set does not read as one person in twelve outfits, and checking every label before it ships.
Independent spec work, made for research and education. Every dish, person and place on this page is generated; none are real. Prices are a September 2026 snapshot.