NutriLog AI Get the app

Guide

Can AI count calories
from a photo?

Short answer: it can tell you what is probably on the plate, and it cannot tell you how much of it there is. That second half is most of the calorie count — which is why the part that matters is not the estimate but whether you can correct it.

Written from a working calorie tracker, with its own measured numbers rather than claims.

What a photo can actually tell a model

Modern vision models are genuinely good at identification. Point one at a plate and it will usually name the dish, list plausible components and attach reference nutrition to each. That part is close to solved, and it is the part demos show you.

The part demos skip is mass. A photograph carries no scale and no density. It cannot show how much oil a curry was finished with, whether that bowl holds 150 g or 300 g of rice, or whether the paneer is the full-fat kind. Those are not edge cases — between them they are most of the number. A dish identified perfectly and portioned wrongly is still wrong by several hundred calories.

So the honest framing is: a photo estimate is a fast first draft. Its value is that it saves you naming eight ingredients from memory. Its danger is that it arrives looking like a measurement.

Why "can you correct it?" is the real question

Every calorie app that scans photos produces roughly the same first number. What separates them is what happens next.

  • The number is final. You get one figure, confidently presented, and no way to argue with it. If it assumed a restaurant portion and you ate half, you are logging a fiction.
  • The number is editable but opaque. You can change the total, but not the reasoning — so you are guessing at a correction to a guess.
  • The ingredients are shown, and each one can be adjusted or dropped. You can see it found naan you did not eat, remove it, and watch the total fall.

Only the third converges on the truth. The first two ask you to trust a number that cannot know the one thing you do know — how much you actually ate.

What people actually do

This is the part that surprised us, so it is worth stating with the number attached. Measured across every meal logged in NutriLog AI on 22 August 2026:

46%of meals were re-logged — something the person had eaten before
8%of meals came from a photo scan

Photo scanning is the feature that sells calorie apps and it is not the feature that gets used. Most days you eat something you have eaten before, and one tap on yesterday's dinner beats any camera, any model and any amount of waiting. If you are choosing a tracker, weight that more heavily than the scan demo.

How to use a photo estimate well

  1. Photograph before you eat, from above, with something of known size in frame if you can. Identification is where the model is strong; give it a clean look.
  2. Read the ingredient list, not the total. The total is a sum of assumptions; the assumptions are where the error lives.
  3. Drop what is not yours and adjust what is. Half a portion is a correction worth 200–300 kcal on a typical restaurant dish.
  4. Once a meal is right, re-log it. The second time you eat it, you are not estimating at all — and that is the path most of your logging will take.

Is it worth using at all?

Yes, for the meals you cannot look up: a mixed plate, a restaurant dish, something a friend cooked. For everything else — packaged food, a dish you eat weekly, anything with a label — a search or a re-log is faster and more accurate. Use the camera where the catalogue cannot help you, and let the catalogue do the rest.

NutriLog AI shows you the ingredients it found in your photo and lets you drop or adjust any of them before it is logged, so your diary records what you corrected rather than what the model guessed. Five AI scans are included to start, and searching, re-logging and the barcode scanner are unlimited and free.

Get it on Google Play

Work out your targets first

An accurate log is only useful against an accurate target. These run entirely in your browser — no sign-up, nothing stored: