Learn/Methodology

Our Methodology: How UPFScore Works

Transparency matters to us. Here's exactly how we estimate the ultra-processed food content of your meals.

Share:

Key Takeaways

  • UPFScore combines AI image recognition with the NOVA food classification system
  • We estimate processing levels based on visible ingredients and typical preparation methods
  • The 1-4 score reflects the weighted average of all identified items in your meal
  • Our approach is designed for practical awareness, not laboratory-level precision

Overview

In short: UPFScore uses AI to identify foods in your photo, then estimates each item's processing level based on the NOVA classification system. Your score reflects the overall ultra-processed content of the meal.

The UPFScore app is designed to give you practical, actionable information about your food choices—not laboratory-precision analysis.

The NOVA Foundation

Our scoring is based on the NOVA food classification system, developed by nutrition researchers at the University of São Paulo, Brazil. NOVA categorizes all foods into four groups:

1Unprocessed or Minimally Processed

Fresh, dried, ground, roasted, boiled, pasteurized, refrigerated, or frozen foods. Examples: fresh fruits, vegetables, eggs, plain milk, fresh meat, fish, nuts.

2Processed Culinary Ingredients

Substances extracted from Group 1 foods, used in cooking. Examples: oils, butter, sugar, salt, flour, pasta.

3Processed Foods

Group 1 foods modified by Group 2 ingredients, using methods like canning, bottling, or fermentation. Examples: canned vegetables, cheese, cured meats, fresh bread.

4Ultra-Processed Foods

Industrial formulations with five or more ingredients, including substances not typically used in cooking (high-fructose corn syrup, hydrogenated oils, protein isolates, additives). Examples: soft drinks, packaged snacks, instant noodles, processed meats.

Our Scoring Process

1. Image Recognition

Our AI analyzes your photo to identify individual food items. It's trained on millions of food images to recognize common dishes, ingredients, and presentations.

2. Classification Mapping

Each identified food is mapped to a NOVA category based on typical versions of that item. A "grilled chicken breast" maps differently than "chicken nuggets."

3. Weighted Scoring

We calculate a weighted average based on the estimated proportion of each item in your meal. A plate that's mostly vegetables with a small processed component will score lower than the reverse.

4. Final Score

The result is a simple 1-4 score reflecting the overall ultra-processed content of your meal. See our score explanation for what each level means.

Key Assumptions

To provide practical estimates, we make certain assumptions:

  • Standard preparations: We assume typical versions of foods unless visual cues suggest otherwise
  • Conservative estimates: When uncertain, we lean toward slightly higher processing estimates
  • Visible content only: We can only score what's visible in the photo—hidden sauces or seasonings may not be fully captured
  • Common ingredients: We assume standard ingredient lists for recognized dishes

Honest Limitations

No estimation system is perfect. Our approach has limitations:

  • We can't distinguish homemade from store-bought versions that look identical
  • Hidden ingredients (inside buns, under sauces) may be estimated rather than observed
  • Regional or unusual dishes may be less accurately classified
  • Photo quality significantly affects recognition accuracy

For a complete discussion, see our accuracy and limitations page.

Frequently Asked Questions

What is the NOVA classification system?

NOVA is a food classification system developed by researchers at the University of São Paulo. It categorizes foods into four groups based on processing level: unprocessed/minimally processed, processed culinary ingredients, processed foods, and ultra-processed foods. UPFScore uses this framework as the foundation for our scoring.

How does the AI identify foods?

Our AI uses computer vision trained on millions of food images to recognize individual items in your meal photo. It identifies foods based on visual characteristics like shape, color, texture, and typical presentation patterns.

Why use AI instead of barcode scanning?

Many meals—especially homemade or restaurant foods—don't have barcodes. Photo-based analysis lets you score any visible meal instantly, without needing packaging information. This makes UPFScore practical for real-world eating situations.

How do you handle uncertainty?

When the AI is uncertain about an item, it uses conservative estimates based on typical versions of that food. We'd rather slightly overestimate processing than give false confidence. The score represents our best estimate given visible information.

Sources & Further Reading

Educational Information Only

This content is for educational and awareness purposes only. It is not medical or dietary advice. Individual situations differ—please consult a healthcare professional for personalized guidance.

Check Your Meal's UPF Score

Ready to see how processed your meals really are?

Curious to see our methodology in action? Scan a meal and see how the scoring works with your own food.

Join the free early access list to try the app as soon as it's ready and get occasional tips on cutting ultra-processed foods. We'll only email you with important updates—no spam, ever.

Get My UPF Score Early Access

About UPFScore

UPFScore is on a mission to help people understand and reduce ultra-processed foods in their diet. Our AI-powered app makes it easy to see how processed your meals really are.

Built by someone passionate about making healthy eating simpler and more accessible for everyone.

Learn more about us and our mission →
Early Bird Offer: 50% Off for Life

Ready to See How Processed Your Plate Is?

Most people are surprised by how high their UPF Score is the first time they scan a meal—join early access to see your own baseline.

Free to join. We'll only email you with important updates and tips on cutting ultra-processed foods. No spam, ever.