World Flavor Atlas
Show HN: World Flavor Atlas. Radzikowski and Chen's food embeddings in 3D
Last verified:
What is World Flavor Atlas?
The World Flavor Atlas is an interactive 3D visualization tool that explores food ingredient embeddings from the Epicure research paper by Jakub Radzikowski and Josef Chen. It visualizes 1,790 canonical food ingredients in a three-dimensional space based on machine learning embeddings that capture culinary relationships including taste, texture, nutrition, geography, and culture.
The tool displays three sibling embedding models side-by-side: Epicure-Cooc (recipe co-occurrence only), Epicure-Core (blended chemistry and recipe context), and Epicure-Chem (chemistry-focused via FlavorDB compound metapaths). Users can explore nearest-neighbor ingredient pairings, observe cuisine macro-region clusters (East Asian, South Asian, Mediterranean, Latin American, Eastern European, Western Atlantic, Southeast Asian, Japanese), and understand how different random walk schemas affect the embedding geometry.
Key features include interactive 3D navigation of ingredient embeddings, UMAP projections colored by cuisine or food group, top-5 nearest neighbor lookups for any ingredient, visualization of 20 interpretable ICA factors per model, and display of 150-200 named culinary modes. The tool helps users understand emergent geometry in food embeddings, including supervised directions for sensory categories and macronutrients.
This tool is designed for computational gastronomy researchers, data scientists working with food embeddings, chefs interested in data-driven ingredient pairing, and anyone studying the intersection of machine learning and culinary science. It serves as an educational resource for understanding how skip-gram embeddings capture culinary knowledge from multilingual recipe corpora.
The visualization was created by the author after reading the Epicure paper to understand it more deeply, making it a companion exploration tool rather than an official release from the paper authors.
World Flavor Atlas pricing
Pricing model: Freemium
Free - The World Flavor Atlas is hosted on GitHub Pages at tomacco.github.io/world-flavor-atlas with no cost, no account required, and no paid tiers. It is a personal exploration visualization created by the author after reading the Epicure research paper.
World Flavor Atlas pros
- Interactive 3D visualization of food ingredient embeddings
- Shows all three Epicure model variants side-by-side for comparison
- Visualizes 1,790 canonical ingredients from multilingual recipe corpus
- Displays cuisine macro-region clusters clearly separated in 3D space
- Shows top-5 nearest neighbor ingredient pairings with cosine similarity scores
- Reveals 20 interpretable ICA factors per embedding model
- Displays 150-200 named culinary modes per model
- Free and open access via GitHub Pages
- Helps understand chemistry-vs-recipe-context spectrum in embeddings
- Visualizes supervised directions for 27 sensory and nutrient probes
- Shows 8 cuisine macro-regions with clear clustering
- Based on 4.14M recipes from 11 sources across 7 languages
- Integrates FlavorDB chemistry data with recipe co-occurrence
- No account or installation required to use
- Educational tool for understanding food embedding geometry
World Flavor Atlas cons
- Created as personal exploration project, not official paper release
- Code and trained Epicure artifacts not released by paper authors
- No search functionality mentioned for finding specific ingredients
- No filtering options for ingredient categories or cuisines
- Limited to viewing pre-computed embeddings, cannot train custom models
- No export functionality for pairing recommendations
- 3D visualization may have performance issues on older devices
- No documentation or tutorial for first-time users mentioned
Frequently asked questions about World Flavor Atlas
What is the World Flavor Atlas?
The World Flavor Atlas is an interactive 3D visualization tool that displays food ingredient embeddings from the Epicure research paper. It visualizes 1,790 canonical food ingredients in three-dimensional space based on machine learning embeddings that capture culinary relationships including taste, texture, nutrition, geography, and culture across three sibling models: Cooc, Core, and Chem.
What are the three Epicure models shown?
Epicure-Cooc walks only the recipe co-occurrence graph with no compound nodes. Epicure-Core blends both chemical and recipe-context signal by injecting pure ingredient-ingredient walks alongside typed-compound metapaths. Epicure-Chem walks only the typed-compound graph with no ingredient-ingredient walks, representing the chemistry extreme of the family.
How many ingredients are in the embedding vocabulary?
The embeddings contain 1,790 canonical ingredients normalized from roughly 200,000 unique ingredient strings using an LLM-augmented pipeline that classifies terms, performs semantic clustering, and includes manual curation. This comes from 4.14M recipes aggregated from 11 sources spanning seven languages.
What cuisines are represented in the visualization?
The tool shows 8 cuisine macro-regions: East Asian (Chinese, Korean), Western Atlantic (American, British, German, Scandinavian), Mediterranean (Italian, French, Iberian, Greek, Levantine, North African, Turkish), Eastern European (Russian, Ukrainian, Polish, Hungarian, Georgian), Southeast Asian (Thai, Vietnamese, Filipino, Indonesian, Malay), South Asian (Indian, Pakistani, Sri Lankan, Bangladeshi), Latin American (Mexican, Caribbean, Brazilian, Peruvian, Colombian), and Japanese.
What is the embedding dimension size?
All three Epicure models use 300-dimensional embeddings. They share identical architecture and hyperparameters including 100 walks per node, walk length of 50, context size of 7, 5 negative samples, batch size of 32,768, learning rate of 0.0025, and 20 epochs of training using SparseAdam optimizer.
How are ingredient pairings calculated?
Ingredient pairings use top-5 cosine nearest neighbors plus the closest emergent mode based on cosine similarity to the mode pole. The tool displays cosine similarity scores for each pairing, with examples like chicken pairing with garlic (0.39), onion (0.37), black pepper (0.36) in Epicure-Cooc.
What is FlavorDB and how is it used?
FlavorDB contains 2,247 typed compound nodes across 15 flavor categories (balsamic, citrus, earthy, fatty, floral, fruity, green, meaty, minty, nutty, spicy, vegetable, wine-like, woody, plus residual). The Core and Chem models use an 80,019-edge typed ingredient-compound graph from FlavorDB to walk compound-mediated metapaths, while Cooc uses only co-occurrence without compounds.
What are the emergent factors and modes?
Each model has 20 stable ICA factors discovered through multi-seed-stable FastICA decomposition on food-group-residualised embeddings. These yield 150-200 named culinary modes per model (Cooc: 150 modes across 41 properties, Core: 193 modes across 44 properties, Chem: 200 modes across 43 properties) with mean coherence 0.611/0.833/0.703 against random-pair baselines of 0.097/0.348/0.115.
Who created the World Flavor Atlas?
The World Flavor Atlas was created by an individual who read the Epicure paper by Jakub Radzikowski and Josef Chen from KAIKAKU.AI and started playing with a visualization to understand the paper more deeply. It is a personal Show HN project, not an official release from the paper authors.
Can I download or use the Epicure models?
No, according to the Epicure paper, the code and trained artifacts are not released at this time. The World Flavor Atlas visualizes the embeddings but the underlying models cannot be downloaded or used independently. This is a limitation noted in the paper's conclusion.