Not very, on most real meals. A photo can't see what's in your food, and the hidden stuff is where the calories are. In published studies, photo-only AI was off by about 30 to 36% on average*. In our own test, it was off by 53%* on a typical plate.
If you log from photos, your day's number may be lower than your real day. Ramen is the iPhone food diary you type into, so the oil and butter get counted and you can see how.
What a photo can't see
A camera sees the top of your food, in the light you had. That's it. Everything else is a guess.
- Oil. It soaks in. A fried egg looks the same with a teaspoon of butter or a tablespoon.
- Butter. Melted into rice or toast, it looks like nothing.
- Dressing and sauce. Tossed through a salad, it vanishes.
- Portion size. From above, a deep bowl and a flat plate look alike.
- What's inside. Wraps, burritos, sandwiches, pies. The filling is a guess.
- What's in the cup. A latte photo can't say oat milk or whole, syrup or none.
Fat has 9 calories a gram. Carbs and protein have 4. So the small stuff you can't see moves the total the most.
What researchers found
- NIH researchers tested four photo calorie apps on 102 meals. All four came in low, by about 250 to 345 calories per meal on average*. Cal AI was the furthest off, at 345 low (source).
- A 2025 study of ChatGPT and another chatbot on 52 photos. They were off by about 36% on average, and they guessed lower and lower as the portions got bigger (source).
- A 2025 study of ChatGPT on 195 dishes. Off by 30.5% from a photo alone. When it also got a list of what was in the food, that dropped to 13.9% (source).
Our test
We gave the same AI 200 real cafeteria plates. Once as a photo, once typed out. On a typical plate it was off by 53%* from the photo and 3%* typed out. Typed was closer on 162 of the 200 plates. A photo guess landed close to the real calories, within 20%, on only about 1 in 5 plates*. Typed out, about 3 in 4 did. Every plate is on our accuracy page.
What it looks like in Ramen
You type: burrito, chips and guac
Ramen shows: a calorie number beside the line, split into the burrito, the chips and the guac, with a note: "Counted 1 large burrito and a handful of chips. Assumed 2 tbsp of guacamole."
Ate half the chips? Type half the chips. The fix is applied, with Undo, and Ramen can remember it for that dish.
A typed line also gets the things a lens skips: the oil in the pan when you write "fried", the sauce you name, the size of the bowl when you say "big bowl", and the latte or handful of nuts you'd have forgotten to snap.
4 quick tips
- Name the fat. "Fried in butter." "About a tablespoon of olive oil." "Creamy sauce." Say if it was fried or baked, too.
- Count the extras. The drink, the bites while you cooked, the handful from someone else's bag.
- Be specific. "Bowl of pasta" gives any AI little to work with. "Two cups of pasta in a creamy sauce" gives it plenty.
- Fix the number when you can see why it's off. An estimate you can read and change beats one you can only accept.
If you already log from photos
You don't have to throw out your history. But for a few days, check one thing: pick a meal that had a sauce, a dressing or a pan of oil, and ask what the camera could not have known. If the answer is "a lot", that day's number is probably low, and a low number on a cut hides the gap between what you think you ate and what you ate.
Ramen is the iPhone food diary that closes that gap with words. It has no camera. Pick lose, hold or build and it sets daily calories, protein and other macros, with a pace that never changes on its own. Each line shows what it counted and what it assumed, so a number you doubt is a number you can change in a sentence. Your weigh-ins then go on a journey chart, which shows whether your logged days and the scale agree. Each week a verdict, three numbers and a few highlights tell you how it went, and a sticker marks the day you reach your goal.
Try Ramen
Monthly costs $4.99. Yearly costs $29.99, under $2.50 a month, with a 3-day trial and a reminder before it ends.
A number you can read, question and fix is worth more than one you can only accept. Type your next meal into Ramen and see what it counted and why. Ramen is coming to iPhone, and the waitlist is how you get in first. Join the waitlist
*Our test: 200 cafeteria plates from Google's Nutrition5k research data (CC BY 4.0), each estimated by the same general AI from a photo and from a typed description. The photo side was a general AI, not any particular app. NIH figures are from a preliminary conference abstract (July 2026, as reported by Healio), and the 30 to 36% range comes from the two peer-reviewed studies above.