A photograph gives software a shape, a surface and whatever else is in frame — nothing more. We ran our own benchmark against a kitchen scale rather than guess at the answer, and it is worth stating plainly before anything else: this is less precise than a scale by a wide margin, and no amount of clever prompting closes that gap. Here is what the numbers actually show.
The typical error is close to half
Across the objects we tested, a single estimate typically misses the true weight by roughly a third to a half. That is not one unlucky model or one bad photograph — it holds across several vision models we tried, on plain single photographs with objects weighed beforehand on a calibrated scale. A grapefruit or a cucumber will usually land in the right neighborhood; landing within ten percent is the exception, not the rule.
It is scatter, not a fixable bias
The first thing anyone suggests on hearing “half the time it is off by a third” is: multiply the estimate by a correction factor and call it fixed. We tested that directly. Most models do lean slightly high, so a correction factor exists in principle. Applying the best possible one removes only five to ten percentage points of error, and the typical miss still sits around half. A consistent nudge cannot fix an error that is mostly random scatter around the true weight rather than a steady lean in one direction.
The same object, a different frame, a different answer
Reshoot the same object under a different condition — a different angle, a different background, a size reference added or removed — and the estimate for that one object can shift by tens of percent, with nothing about the object itself having changed. A dinner plate or a ceramic mug photographed twice can come back at two quite different weights. The camera did not change; the light and framing did, and that alone moves the number.
A size reference is not a reliable fix
Our guide to photographing an object suggests laying a coin or a battery beside it, and that is worth trying — it gives the software a known size to work against, and for at least one model we tested it helped substantially. But it is not a guarantee, and our benchmark is the reason to say so plainly. Adding a reference helped some models and made others noticeably worse. In one run, a ceramic mug weighed at 264 grams on a real scale was estimated at 10 grams once an AA battery was placed in the same frame — worse than leaving the battery out entirely. A reference only helps if the model reads it correctly; misread, it hands the model a false sense of scale, and the result is worse than having no reference at all. Try it, take the shot both ways if you can, and do not treat either version as automatically the better answer.
What it is good for
None of this makes the method useless — it makes it a particular kind of tool. It is good at telling a heavy object from a light one, at putting a number on something when no scale is nearby, at giving a rough order of magnitude, and at comparing two objects side by side. It is not good for anything where the number has to be right: postage that is charged by weight, a dose, a sale priced by weight, or anything with legal or commercial weight on the line. For those, use a scale. For “roughly how much does this weigh,” a photo is faster than finding one.
The scale of what we tested
This is based on our own testing, not a general claim about all photo-based weight tools: on the order of twenty physical objects, several vision models, single photographs, each object weighed in advance on a calibrated scale. It is a real benchmark, not a marketing figure, and it is small enough that we describe the results in ranges and tendencies rather than precise percentages we cannot stand behind at a larger scale.