How the pairing works
Every time you press Build, the generator puts together about sixty complete pizzas, scores each one as a whole, and shows you one of the best. A pizza scores well when its ingredients pair well, its tastes balance, and its crust can carry the load.
1. Do these two ingredients go together?
Every pair of ingredients gets three independent opinions.
Chef knowledge. Each of our 234 ingredients lists the pairings pizzaiolos and pizza cookbooks rely on, plus the few combinations to avoid.
Recipes. Using FlavorGraph's analysis of about a million recipes, we measure how much more often two ingredients appear together than chance would predict (normalized pointwise mutual information). Basil and tomato score near the top. Pairs in the top quarter earn a green proven classic dot.
Chemistry. FlavorGraph also links ingredients to the flavor compounds they contain. Ingredients that share many aroma compounds tend to taste related: this is the food-pairing hypothesis behind combinations like strawberry with basil. 215 of our ingredients have compound data.
When a pair shares a lot of chemistry but rarely shows up in recipes, we call it a surprise, marked with a yellow star. The adventure slider decides how much weight surprises get: at Classic the generator trusts recipes; at Adventurous it trusts chemistry.
2. Does the whole pizza balance?
Pairs aren't enough: a pie of five individually-great matches can still be all fat and no acid. A taste model adds up salt, richness, acidity, sweetness, heat, savoriness, bitterness, freshness and smoke across the ingredients, weighted by how much of each you'd use. It rewards the contrasts good pizzas lean on (richness cut by acid, salt met by something sweet, heat calmed by something creamy) and flags pies that are all one note, under-seasoned or too sweet.
3. Can the crust carry it?
Ingredients carry a weight from feather-light (lemon zest) to heavy and wet (sausage, squash). Each of the 11 crust styles has a load budget: a 90-second Neapolitan takes far less than a Detroit pan pie. The generator also avoids piling up proteins and picking two ingredients that do the same job, such as two similar salamis.
The crust style is chosen last, from the ingredients: Detroit for brick cheese and cupping pepperoni, New Haven for clams and garlic, Flammkuchen for crème fraîche and onions. You can fix a style yourself under Diet, crust style and more.
Where the data comes from
Recipe and compound data: Donghyeon Park, Keonwoo Kim, Seoyoon Kim, Michael Spranger and Jaewoo Kang, “FlavorGraph: a large-scale food-chemical graph for generating food representations and recommending food pairings,” Scientific Reports 11, 931 (2021), doi:10.1038/s41598-020-79422-8. The graph combines Recipe1M with FlavorDB compound data and is published under the Apache 2.0 license at github.com/lamypark/FlavorGraph. We map each of our ingredients to its closest FlavorGraph ingredients and precompute 27,261 pair scores.
Inspiration for how we use the data: the FlavorGraph Playground, which explores the same graph in the browser and puts its three signals side by side: cooked together in recipes, shared flavor compounds, and similar culinary role.
Pizza names come from a small language model when it's available, and from a simple naming rule when it isn't. The model only names the pizza; it never chooses the ingredients.