How to estimate portions from a photo (and fix the AI's guess)

Many calorie-tracking errors aren’t about what you ate. They’re about how much. Recognizing “rice with chicken and broccoli” is usually the easy part, for you and for an AI; telling 150 grams of rice from 250 grams in a photo is much harder.

The good news is that portion estimates get better with a few habits: how you frame the photo, what you put next to the plate, and a handful of benchmarks you know by heart. Here’s what works, what the research says, and how to correct the estimate once it’s in your log.

Why portions are so hard to judge

Three things work against you.

Large portions shrink in your mind. In a classic study of more than 7,000 portion judgments, people tended to underestimate large portions more than small or medium ones.

Photos have no built-in scale. Without something of known size in the frame, a small plate up close and a big plate further away can look the same. Researchers working on automatic food measurement compare the reference object to the scale bar on a map: without one, sizes can’t be determined.

Volume is slippery. A cup isn’t always a cup. USDA FoodData Central lists one cup of cooked spaghetti, loosely filled, at 124 grams and 196 calories; packed down, the same cup holds 151 grams and 239 calories. That’s about a fifth more food from the same measure.

How to take a photo that’s easy to estimate

Whether a person or an AI reads your photo, these habits give it more to work with.

Shoot before the first bite

Research tools built around food photos, like the mobile food record, ask people to photograph meals before eating and again afterward. In an early validation study, portion estimates from digital photos of meals and their leftovers were comparable to estimates made by observers looking at the food in person. The habit is worth borrowing: if you don’t finish, a second photo or a quick portion adjustment keeps your log honest.

Tilt to about 45 degrees

Straight-down shots hide height. In a 2025 study, adults matched real meals to photos taken from different angles. For solid foods, 45-degree photos led to the most accurate matches. Overhead shots led people to underestimate rice, because its height in the bowl was hard to see from above. For drinks, a steeper, more side-on angle worked best, and combining angles improved accuracy for most foods.

So hold your phone about halfway between straight down and side-on. For a drink, go lower and more side-on so the fill level shows.

Include something of known size

A fork, a standard soda can or your phone all give a sense of scale. Your own plates are the easiest reference of all: measure their diameter once and “half the plate” becomes meaningful. Researchers have used a circular plate of known size as the scale reference for food photos when no special marker was available.

Get the light right

Bright, even light, such as a nearby window, shows edges, textures and sauces clearly. Avoid strong shadows, colored lighting and flash glare when you can. Image researchers note that changes in illumination and viewpoint alter how foods look, which makes them harder to identify.

Show what’s hidden

If food is stacked, spread it out a little. Keep sauces and dressings visible, or mention them in a note. Fat that has soaked in or melted is one of the hardest things to estimate from any photo, and it’s a big reason photo apps tend to underestimate calories.

Hand-based estimates: handy, but rough

Using your hand as a measuring tool is appealing because it’s always with you. The research is mixed.

In a 2016 study, 67 adults estimated 42 pre-weighed foods and drinks using hand-based methods and household measures like cups and spoons:

• Using finger widths as a ruler to measure the length, width and height of regularly shaped foods worked reasonably well. For 80 percent of those items, the typical estimate was within 25 percent of the true weight, compared with 29 percent using cups and spoons.
• Irregular cuts like chicken breasts, fish fillets and steaks were overestimated with the finger method.
• Using a fist to size up amorphous foods, such as rice or cereal, worked poorly: for most of these foods, estimates were off by more than 50 percent.
• Thumb and fingertip estimates for spreads were also significantly off.

So treat your hand as a sanity check, not a scale. It’s useful for noticing when a portion is clearly bigger than your usual one, and least reliable for calorie-dense foods like oil, nut butter and cheese, where a small misjudgment carries a lot of calories.

Portion benchmarks worth knowing

These USDA FoodData Central values make handy anchors. They’re for plain foods, so cooking fats, sauces and toppings are extra.

A few patterns stand out:

• Shape changes how much fits. The same cup holds noticeably fewer grams of penne than of spaghetti.
• Fats are dense. One tablespoon of olive oil has more calories than half a cup of cooked rice.
• Cooked and dry weights are different foods. USDA lists raw white rice at 365 calories per 100 grams versus 130 for cooked, so weighing cooked rice but logging it as raw nearly triples the number. Match the entry to the state you weighed.

Calibrate your eye with a scale

You don’t need a kitchen scale for everyday logging, but a short calibration period pays off. For a week or two, weigh the foods you eat most often in the dishes you actually use: your usual scoop of rice, your bowl of cereal, your pour of oil, your slice of cheese. Photograph the same portions as you go.

You’ll quickly learn what 150 grams of rice looks like on your own plate, and your corrections to an AI estimate will get faster and more confident. Repeat now and then, especially if you get new bowls or plates.

From photo to log: review and correct

A photo estimate is a first draft. In Calorix, every AI scan opens a review card listing each ingredient with its portion, grams, calories and a confidence badge. A quick routine:

1. Start with the lowest-confidence items. Those are the parts of the estimate the AI is least sure about.
2. Check the grams on calorie-dense ingredients. Oils, cheese, nuts, sauces and meats move the total the most.
3. Add what the photo can’t see. If it was cooked in butter or dressed at the table, add it as its own ingredient, or tell Fix with AI by text or voice, for example “add 1 tablespoon olive oil”. You’ll see the proposed changes before you accept them.
4. Match the amount you actually ate. Use the portion slider (½, 1, 1½ or 2), or edit the grams directly if you weighed something.
5. Add a note and re-analyze when the picture misleads. A note like “half portion” or “fried”, followed by a re-analysis of the stored photo, gives the AI the context it was missing.

Totals always recalculate from the ingredients, so every correction flows into your macro tracking. If you override a value by hand, Calorix keeps the AI’s original so you can revert to it. For anything with a label, barcode or nutrition-label scanning skips the guesswork entirely.

The bottom line

Better portion estimates start with better inputs. Shoot before you eat, at about 45 degrees, in good light, with something of known size in the frame. Use your hand for rough checks, keep a few USDA benchmarks in mind, and weigh your most common foods once in a while to stay calibrated.

Then take ten seconds to review the AI’s estimate, especially the fats and the starchy sides. That review is the part a camera can’t do for you, and it’s what turns a quick photo into a log you can trust. If you’re setting targets to log against, the macro calculator is a good next step.