How to photograph food for calorie counting

The camera can see identity and surface area. It cannot see oil, density or what is underneath. Photograph for the first two and tell the app the third.

What a photograph actually contains

It helps to be precise about what any image-based estimator, human or machine, is working from. A photo carries good information about what a food is and about its visible area. It carries weak information about volume (height is foreshortened, and what is underneath is inferred) and essentially none about composition — whether the yoghurt is full-fat, how much oil went into the pan, what the sauce is made of.

The Nutrients study that measured this escalated the context given to the model in steps. Photo alone: 30.5% mean error on energy. Photo plus non-visual context such as the cooking method: 24.4%. Photo plus a list of ingredients: 13.9%.¹ Fat error fell from 8.7 g to 4.3 g across the same steps. The rules below are ordered by how much of that gradient each one recovers.

Rule 1 — Add one sentence about the fat

This is worth more than every other rule combined. Oil, butter and dressings are the single largest source of error in photo estimation because they are invisible once absorbed and calorie-dense: a tablespoon of oil is about 120 kcal, and a plate of roasted vegetables can hold two or three. "Roasted in a tablespoon of olive oil", "dressing is two tablespoons of ranch", "cooked in butter" — one phrase recovers the calories the image lost.

If the app allows text or voice alongside the photo, use it every time fat was involved. If it does not, correct the fat line item by hand before saving.

Rule 2 — Shoot from 45 degrees, not straight down

Overhead shots show area beautifully and height not at all: a shallow bowl of rice and a heaped one look identical. Side-on shots show height and hide area. About 45 degrees shows both, which is why it is the angle most food-recognition datasets are built around. The Nutrition5k dataset that Google's researchers used to train and evaluate their estimator captured dishes from multiple angles and with depth sensing precisely because a single overhead frame under-determines volume.²

Rule 3 — Keep something of known size in the frame

Without a scale reference, a large portion photographed from far away and a small portion photographed close up can produce the same image. A fork, a standard dinner plate, your hand, or the edge of your phone gives the estimator a ruler. The plate itself is the best reference of all, so keep its whole rim in the frame rather than cropping to the food.

Rule 4 — Photograph before the first bite

Obvious, and routinely skipped. A half-eaten plate is a guess about what the full one was, and the foods that vanish first (fries, bread, the crispy bits) are usually the dense ones. If you forget, photograph what is left and log the difference, or photograph the empty plate and describe what was on it — a described meal still beats a forgotten one.

Rule 5 — Separate what hides other things

Estimators struggle most with homogenised and layered dishes: curries, stews, casseroles, smoothies, anything under a blanket of cheese.¹ Where you can, plate so that components are visible — sauce beside the rice rather than over it, toppings spread rather than piled. Where you cannot, do not fight it: name the components instead. "Chicken curry, about a cup of sauce, a cup of rice" is Rule 1 applied to occlusion.

Rule 6 — Photograph the label, not the food

For anything that came in a package, the manufacturer's label is strictly better information than any image, because it is a measurement rather than an estimate. Photograph the nutrition panel and the serving size, and log by weight or by number of servings. This is the one case where the food photo is the worse option.

Rule 7 — Good light, no filters, no flash

Colour carries identity: the difference between brown rice and white, between a lean and a marbled cut, between a full-fat and a light dressing is often colour and sheen. Daylight or neutral indoor light, no colour filters, and no direct flash (which flattens texture and blows out highlights). Hold still; motion blur removes the fine texture that separates, say, couscous from quinoa.

What still will not work

None of this makes a photograph a scale. Expect the estimate to remain rough for drinks in opaque cups, for dishes where the calories are in an invisible sugar syrup, and for anything where the portion is mostly below the surface. The honest way to use a photo app is to treat its number as a first draft, check the one or two items that dominate the calories, and let the rest go — the two-week weight trend, not the per-meal estimate, is what you are actually managing.

Where Wellix fits

Wellix is built around Rule 1. Every scan can carry a text or voice note, the result is an itemised list you review before saving, and when an item is genuinely ambiguous — sauce or no sauce, cooked in oil or dry — Wellix asks you one clarifying question rather than guessing. The photo does what a photo can do; the sentence does the rest. How much to trust the result, in numbers, is laid out in how accurate are AI calorie counting apps.

Frequently asked questions

What is the best angle to photograph food for a calorie app?

About 45 degrees above the plate, so the image shows both the area the food covers and its height. Straight-down shots hide height, side-on shots hide area, and both make portion volume a guess. Keep the whole plate rim in the frame.

Does adding a description improve an AI calorie estimate?

Yes, more than any camera technique. In a peer-reviewed evaluation on real dishes, energy error fell from 30.5% with a photo alone to 13.9% when the ingredients were listed. The most valuable single detail is the cooking fat, because oil and butter are invisible once absorbed and calorie-dense.

Why do AI calorie counters get oil and dressings wrong?

Because they cannot be seen. A tablespoon of oil absorbed into vegetables changes the image not at all and the calories by about 120 kcal. Tell the app how much fat was used, or correct the fat item before saving.

Should I photograph packaged food?

Photograph the nutrition label and the serving size instead. The label is the manufacturer's measurement, which is strictly better than any estimate from an image of the food.

Can I log a meal I forgot to photograph?

Yes. Photograph what is left and log the difference, or describe the meal in text or voice. A described meal is far better than a skipped one: the meals people forget to log are systematically the larger ones.

References

1. Rodríguez-Jiménez M, Martín-del-Campo-Becerra GD, Sumalla-Cano S, Crespo-Álvarez J, Elio I. Image-Based Dietary Energy and Macronutrients Estimation with ChatGPT-5: Cross-Source Evaluation Across Escalating Context Scenarios. Nutrients. 2025;17(22):3613. doi:10.3390/nu17223613. pmc.ncbi.nlm.nih.gov/articles/PMC12655113
2. Thames Q, Karpur A, Norris W, Xia F, Panait L, Weyand T, Sim J. Nutrition5k: Towards Automatic Nutritional Understanding of Generic Food. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021, pp. 8903–8911. arxiv.org/abs/2103.03375

Wellix provides general nutrition information and is not medical advice. It does not diagnose, treat, or prevent any condition. Calorie and macronutrient targets are estimates produced by predictive equations, and individual requirements vary. Consult a qualified healthcare professional before making significant dietary changes, particularly if you are pregnant or breastfeeding, or have a history of disordered eating, diabetes, kidney disease, thyroid disease, or any other medical condition.

Related

• How accurate are AI calorie counting apps?
• AI calorie counter vs a food scale: which should you use?
• How to track calories at a restaurant
• Calories in foods: per 100 g and per serving

Photograph it, then say the one thing it cannot see

Wellix takes a photo plus a sentence, shows you every item before saving, and asks when it is unsure instead of guessing.