Bo guide
Short answer
Include a size reference, shoot at a 45-degree angle rather than overhead, separate overlapping foods, photograph before eating, and correct the estimate when you know what went into the cooking.
General information, not medical advice. Written from published sources, which are listed at the end. It is not a substitute for a doctor, pharmacist or registered dietitian who knows your situation.
In this guide11 sections · 4 min
- What the model is actually trying to do
- 1. Give it something to measure against
- 2. Shoot at an angle, not straight down
- 3. Separate overlapping food
- 4. Photograph before you eat
- 5. Correct what you actually know
- Lighting, and why it matters more than you would think
- Which foods photograph well
- Photograph the packet, not the plate
- Consistency beats accuracy
- When not to bother with a photo
- Common questions
Photo calorie logging is only as good as the photo. These habits reportedly improve accuracy by 10 to 15 percentage points — a bigger gain than switching apps.
What the model is actually trying to do
Worth understanding, because it tells you where to help. From one flat image the app has to solve three separate problems: identify each food, estimate its volume, and infer its density and preparation.
Identification is the part these systems are good at. Volume is much harder from a single viewpoint with no depth information. Preparation is close to impossible — grilled and pan-fried chicken look nearly identical, and the difference is a hundred calories of oil. Every habit below is aimed at the second and third problems, because the first is already handled.
1. Give it something to measure against
Scale is the hardest thing to infer from an image. Include a known object: your hand, a fork, a standard dinner plate, a can. Without a reference, the model is guessing portion size from context alone, and that is where the largest errors start.
2. Shoot at an angle, not straight down
Overhead photos flatten everything and hide depth. A bowl of pasta shot from directly above looks identical whether it is two inches or five inches deep. A roughly 45-degree angle shows both the surface area and the height.
3. Separate overlapping food
Push components apart before shooting. Rice hidden under curry, or chicken sitting on potatoes, simply does not register. Ten seconds of rearranging is worth more than any model improvement.
4. Photograph before you eat
Obvious, easy to forget. A half-eaten plate is hard to estimate from, and reconstructing it later is guesswork. If you forget, log it by description instead — voice logging is usually more accurate than photographing leftovers.
5. Correct what you actually know
If you cooked the meal, you know things the camera cannot: the two tablespoons of olive oil, the butter in the sauce. Editing the estimate takes seconds and is the highest-quality data the app will ever receive.
Lighting, and why it matters more than you would think
Recognition depends on the model seeing colour and texture accurately, and most kitchen and restaurant lighting distorts both.
- Daylight is best, from the side rather than behind you.
- Avoid your own shadow falling across the plate — the commonest problem when shooting from directly above.
- Warm restaurant lighting pushes everything orange, which makes sauces and dressings harder to identify.
- Phone flash flattens texture, which is exactly the cue that distinguishes, say, grilled from fried.
Which foods photograph well
| Photographs well | Photographs badly |
|---|---|
| Plated meat and vegetables | Soups and stews |
| Whole fruit | Smoothies and drinks |
| A sandwich, opened up | Anything blended or uniform |
| Distinct components, spaced out | Nuts, seeds, loose handfuls |
| Grilled or roasted food | Anything with hidden sauce or dressing |
The pattern is that photos work when the components are distinguishable and their volume is readable. As soon as food becomes a uniform mass, a photo carries almost no information a description would not carry better.
Photograph the packet, not the plate
For anything packaged, the barcode or nutrition label is far more accurate than any photo of the food itself, because it is a lookup rather than an estimate. Cereal, yoghurt, ready meals, protein bars, bread — scan or type these rather than photographing them.
Reserve photo logging for cooked and restaurant food, where no label exists. That is the case it is actually good at.
Consistency beats accuracy
The most useful reframing, and the one that takes the pressure off. If your logging is consistently 15% low, your daily numbers are wrong but your trend is right — and the trend is what you act on. You find your real maintenance empirically, in the units your app reports, and the bias cancels out.
What ruins that is inconsistency: photographing carefully on weekdays and guessing at weekends, or logging thoroughly when the week is going well. A steady, honest, slightly-wrong log is far more useful than an erratic accurate one. This is also why measuring your own maintenance from your own logs beats any calculator — it is calibrated in the same units as your errors.
When not to bother with a photo
Some foods photograph badly and are faster to describe: drinks, soups, handfuls of nuts, anything mixed into a uniform mass. See also where photo estimates go wrong. For those, voice or text logging is both quicker and more accurate.
