To get the best results from AI food recognition, shoot from directly above in good light with every food visible, add a quick note about oil, sauces or anything hidden, and check the portion before you save. Portion size causes most of the error. In the Nutrition5k study, calorie estimates from a single food photo were off by about 26% on average, compared with about 9.5% when the model knew how much the food weighed.
None of this takes more than a few seconds per meal.
Key takeaways
- Angle: directly overhead, whole plate in frame. Add a side shot for tall food.
- Light: daylight or bright, even light. No dim corners, no harsh flash glare.
- Visibility: if the camera can't see it, the AI can't count it. Spread food out and lift what's buried.
- Add-ons: log oil, butter, dressing and sauces yourself. They're the most common misses.
- Portion: always glance at the estimated amount and fix it. That single check matters most.
Why photo quality and portion size matter
A food recognition model has two jobs: work out what's on the plate, and how much of it there is. The first is the easier part. Judging amount from a flat picture is the hard part, and it's where most calorie error comes from.
Nutrition5k, a Google research dataset of 5,006 cafeteria plates with weighed ingredients, shows the gap clearly:
| What the model had to go on | Average calorie error |
|---|---|
| The food's weight was known | 9.5% |
| Photo plus 3D depth data | 16.5% to 18.8% |
| One ordinary photo | 26.1% |
People do no better by eye. In the same study, untrained people misjudged the weight of food on simple plates by 53% on average, and nutritionists by 41%.
The tips below do two things: give the AI a clear view, and help you catch the portion errors it will still make. For the technical side, see how AI handles portion size estimation.
9 tips for better AI food recognition
1. Shoot from directly above
Hold the phone flat over the center of the plate, lens pointing straight down. An overhead shot shows every food at once, with nothing hiding behind anything else, and keeps sizes in proportion. Shots from table level hide the back of the plate and make near items look bigger.
2. Keep the whole plate in the frame
Step back far enough to fit the entire plate or bowl, including the rim. The plate edge gives a sense of scale. Cropping tightly on the food removes that cue.
3. Add a second angle for tall food
A burger, a layered sandwich, a stack of pancakes or a deep bowl looks flat from above. Take a second photo from about 45 degrees so the height shows.
4. Use good, even light
Daylight near a window is best. Bright overhead kitchen light is next. In a dim restaurant, use your phone's night mode or move the plate toward the light rather than firing the flash, which can bounce glare off sauces, wet salads and white plates. Before you submit, check that you can make out every food in the preview.
5. Make every food visible
If something is hidden, the AI can't count it. Spread items out a little so they don't overlap. Lift noodles out of the broth with chopsticks, open a burrito or wrap halfway, and push lettuce aside so the toppings show. If you're about to mix a grain bowl, take the photo first.
6. Add a quick note
Some apps let you add text or voice context before the photo is analyzed. Use it for what a camera can't see: "cooked in 1 tbsp olive oil", "made with whole milk", "no sugar in the coffee", "half portion of rice". One short line often fixes more than a perfect photo.
7. Log fats and sauces yourself
Oil soaked into vegetables, butter melted into potatoes and dressing tossed through a salad are close to invisible in a photo, and they're calorie-dense.
| Add-on | Amount | Calories (kcal) |
|---|---|---|
| Olive oil | 1 tbsp (13.5 g) | 119 |
| Butter | 1 tbsp (14.2 g) | 102 |
| Mayonnaise | 1 tbsp (13.8 g) | 94 |
| Ranch dressing | 2 tbsp (30 g) | 129 |
Source: USDA FoodData Central. Two tablespoons of olive oil in a home stir-fry add 239 calories that a photo of the finished dish may not show. For lighter options, see our low-calorie sauces and condiments guide.
8. Check the portion before you save
This is the step that matters most. Look at the amount the app estimated (grams, cups or pieces) and ask whether it matches your plate. If you eat the same foods often, weigh them once: knowing that your usual bowl of rice is about 1.5 cups makes every future correction quick.
9. Photograph before you eat
Take the photo before the first bite, while everything is still on the plate. If you don't finish, adjust the portion to what you actually ate rather than taking a second photo of the leftovers.
How to photograph hard foods
| Food | What goes wrong | What to do |
|---|---|---|
| Soups and stews | Depth is invisible; solids sink | Shoot from above, then set the portion by bowl size |
| Noodle and grain bowls | Base is buried under toppings | Lift some noodles or rice into view; note the bowl size |
| Sandwiches, burgers, wraps | Fillings are hidden | Open it or cut it in half; add a side shot |
| Smoothies and coffee drinks | Ingredients can't be seen | Type or say the ingredients and the cup size |
| Casseroles, curries, lasagna | Mixed ingredients | Log as a recipe if you made it, or describe the main parts |
| Salads | Dressing and oil disappear | Log the dressing separately |
Mixed dishes are where a photo struggles most. If you cooked it, logging the recipe once and then the portion you ate beats any photo. Our guide to tracking mixed meals at home with AI walks through it.
When to correct the AI
Review every estimate for a few seconds before saving. Correct it when:
- It named the wrong food, like chicken that's actually tofu.
- It missed something, like nuts on a salad or cheese inside a wrap.
- It got the cooking method wrong, like fried when you baked it.
- The portion looks bigger or smaller than what's on your plate.
- You added oil, butter or sauce it couldn't see.
A corrected estimate is still faster than logging everything by hand, and a rough log beats no log. To see how accurate the AI is before and after corrections, read how accurate is AI calorie counting. For packaged foods, a barcode scan is usually more accurate than a photo; see barcode scanning vs AI photo recognition.
Using these tips in Nutrify
We make Nutrify, so here's how the tips map to it. After you take a photo, or pick one from your gallery, you can add a typed or spoken note before the analysis runs. On the result screen you can change the food name, the portion and each ingredient before you add it to your day. For packaged food, the same camera scans barcodes, and you can also log by Siri, a typed description or a saved meal. Nutrify has a 7-day free trial, then a paid subscription.
Sources
- Thames Q, et al. Nutrition5k: towards automatic nutritional understanding of generic food. CVPR 2021.
- USDA FoodData Central: olive oil, FDC 171413; butter, salted, FDC 173410; mayonnaise, FDC 171009; ranch dressing, FDC 173592.
- Apple App Store, Nutrify - AI Calorie Tracker: logging methods and subscription terms. Checked October 26, 2026.



