Take two photos of the same dish and AI may return different calorie estimates because it interprets clues in two images rather than directly measuring the food on the plate. Small changes in distance, angle, lighting, obstruction, and container alter the visible area and height. Even when the photos look similar, the model’s inferences about the dish and portion may vary.
In nosh, calories come from a chain: identifying the food, estimating its portion, and matching nutrition data. A change anywhere earlier in the chain can change the final number. Different results do not necessarily mean one is entirely wrong. It is more useful to find where the difference began.
First ask whether it is the same plate photographed twice or the same dish eaten twice
Imagine photographing one plate of tomato and scrambled egg once from above while standing and again at an angle while seated. From above, egg covers a large area. At an angle, tomato in front may hide part of the egg and the plate looks narrower. The food did not change, but the shapes presented to AI did.
Another “same dish” may only share a name. Today’s tomato and egg may use three eggs; tomorrow’s cafeteria version may have more tomato, oil, or sugar and a different portion. Here, different calories do not come only from recognition variation. The two meals genuinely have different recipes.
Mixing these situations leads to the wrong conclusion. Differences between two photos of one plate mainly come from photography and inference. Differences between two servings with the same name also include real preparation and portion changes.
Size in a photo is not size in the real world
Move a phone closer and half a bowl of rice can fill the frame. Move it back and the same rice looks small. An overhead image shows how widely food spreads but hides its height. An angled photo reveals height while allowing foods to cover one another.
Containers create illusions too. The same serving fills a small bowl but occupies only the center of a large plate. AI can use a plate, bowl, or utensil as a clue, but without its real dimensions, the object remains a reference rather than a ruler.
The second photo is not simply a repeat of the first. Once camera position or composition changes, the system receives new evidence.
A small change in the name can create a large change in the number
The same chicken may be identified as grilled chicken, skin-on fried chicken, or poached chicken breast, each matching different nutrition data. Even with a similar name, moving a portion from small to medium can amplify the final estimate.
These differences matter more than a few calories of fluctuation. Results of 618 and 636 calories may not justify repeated photos. One result calling steamed fish fried, or half a bowl of rice a full bowl, deserves correction.
The size of the numerical difference is not the only question. Ask whether it changes how you understand the meal.
Do not keep photographing until you get an answer you like
If three photos of one meal produce 580, 640, and 690 calories, it is tempting to save the most appealing result. That does not resolve the variation; it turns the estimate into a draw from a hat.
Keep the clearest image of the whole meal and check three things: Are the main foods correct? Is the amount you actually ate reasonable? Are any drinks, sides, or obvious sauces missing? Correct what you know and leave the rest as an estimate.
nosh currently lets you edit names, calories, portions, and log times. Fixing an error that changes the conclusion is more useful than repeatedly taking photos in search of an ideal number.
For long-term comparisons, keep the method reasonably consistent
If you often log the same breakfast in nosh, keep the distance, angle, and container roughly consistent. Include the full plate, photograph before eating, and use the same rule for whether the drink appears. This does not turn estimates into measurements, but it reduces variation caused by the way you photograph.
Consistency does not require recreating an identical composition every day. What matters is the logging rule: whether the image shows the serving or the amount consumed, whether drinks count, and whether obvious mistakes are corrected each time.
When AI gives two different answers, the best first response is not “take another photo.” Ask what changed in the image. Finding the layer where variation began is more useful than demanding exactly the same number every time.