Computer vision is opening a more accessible path to body composition tracking. This narrative review explains three relevant studies, what they found and what those findings cannot prove about BodyFat AI.
Evidence reviewed September 2026Narrative review3 primary studies
Important disclosure
The studies below evaluated separate computer vision systems. They did not test, validate or establish the accuracy of BodyFat AI. BodyFat AI results are wellness estimates and are not medical diagnoses.
The useful question is not whether every method agrees.
The studies answer a narrower question: how closely each tested method agreed with DXA under that study's conditions. Whether a method accurately follows change over time is a separate question, and the Nana study specifically called for longitudinal research.
Dual energy X ray absorptiometry, known as DXA, was the reference in each study summarized here. The image systems were compared with DXA, and some studies also compared specific bioelectrical impedance devices.
These studies support the broader promise of image based body composition estimation. They do not make every image model interchangeable, and they do not remove uncertainty from an individual result.
STUDY 1
Ferreira and colleagues, 2025
Researchers studied 1,273 adults and compared an artificial intelligence method using two dimensional photos with DXA. They also evaluated skinfolds, ultrasound and several bioelectrical impedance devices.
1,273ADULTS0.98PHOTO CCC IN THIS STUDYDXAREFERENCE METHOD
The photo method showed the strongest reported agreement with DXA among the methods examined. Overall concordance was 0.98. Sex specific concordance was 0.98 for men and 0.96 for women. In the same overall sample, InBody 270 and Omron HBF 514 reported 0.92 and 0.91.
Interpretation still requires care. Performance varied across sex and body mass index groups, measurements followed a study protocol, and the evaluated system was not BodyFat AI.
STUDY 2
Nana and colleagues, 2022
This study compared body composition estimates from two dimensional smartphone images and a Tanita BC 313 foot to foot bioelectrical impedance scale with DXA in 929 adults.
929ADULTS0.90PHOTO CCC FOR MEN AND WOMEN2.8 to 2.9PHOTO RMSE POINTS
For body fat percentage, the image method had a concordance correlation coefficient of 0.90 for both men and women. Root mean square error was 2.9 percentage points for men and 2.8 for women. The tested scale reported lower agreement and higher error in the same sample.
The authors described the method as a promising alternative for home assessment. The paper listed affiliations with Body Composition Technologies and Advanced Human Imaging for several authors. It also called for future research on tracking change over time, so longitudinal performance remained unresolved.
STUDY 3
Majmudar and colleagues, 2022
Researchers evaluated a smartphone camera method called visual body composition in 134 adults. DXA served as the reference, with three consumer bioelectrical impedance scales, two professional systems and air displacement plethysmography included for comparison.
The image method produced the lowest mean absolute error of the evaluated methods at 2.16 percentage points and an overall concordance of 0.96 with DXA. The three consumer scales in that study produced mean absolute errors of 4.48, 4.91 and 5.85 percentage points.
The sample was modest, four participants were excluded for poor image quality, and controlled instructions still mattered. Amazon employees worked for the study sponsor and contributed to study design, analysis and publication. The paper states that the sponsor remained blinded to study data until enrolment was complete and the image estimates had been shared with investigators.
HOW TO READ THE EVIDENCE
Promising does not mean perfect.
What the studies support
In the evaluated systems and study samples, smartphone image methods showed meaningful agreement with DXA and reported stronger agreement or lower error than the specific impedance devices tested in the cited comparisons.
What the studies do not support
They do not prove that all photo methods perform equally, guarantee an individual estimate, replace clinical measurement or validate BodyFat AI.
Method and scope
This is a narrative review of selected studies directly relevant to smartphone image based body fat estimation. It is not a systematic review or a meta analysis, and no results were pooled. Sources were selected for direct comparison with a criterion or reference method and for clear reporting of body fat performance.
Practical takeaway
Use any consumer body composition number as an estimate. Scan under similar conditions, look for direction across repeated measurements and consider the result alongside nutrition, strength, recovery, weight and progress photos.
Use the scan for direction.
Keep every estimate in context and every next step clear.
The male range of 5% to 25% and female range of 10% to 35% are editorial exploration bounds for adults. Very lean endpoints are not recommended everyday targets. They are not clinically validated healthy ranges or personal recommendations. Defaults of 22% and 30% do not identify an ideal percentage.
Each subject uses six individually generated images at 1024 by 1536 pixels. These replace the smaller figures in the earlier image sheets. Their height and centre are aligned before a 2D blend updates at each whole percentage selection. The 47 selectable values are not 47 measured photographs. Intermediate appearances are interpolated illustrations, not measured one percentage point changes. These fictional people have not had body composition measured, and their appearances cannot substantiate an exact percentage. This is a 2D illustration, not a scan, measurement or prediction.
Gallagher and colleagues, 2000 described provisional body fat ranges derived from BMI. Age, sex and ethnic group affected the relationship. This research does not validate the selector endpoints.
We reviewed the participant photographs in Hulmi and colleagues, 2017. In the same female fitness group, average estimates at the competition assessment differed between DXA and skinfold methods. These study photographs illustrate variation. They do not calibrate our fictional subjects, and a group average cannot be assigned to an individual photograph.
Men and women also differ in typical fat distribution. Bredella, 2017 describes sex differences in lean and fat mass and where fat is stored. This informs visual direction, not numerical calibration.
People can look different at the same percentage. Research on regional fat distribution helps explain why a total percentage cannot predict an individual appearance.
Your starting assessment matters. The NIDDK Body Weight Planner, for example, needs starting measurements and activity, alongside a goal and timeframe. A selected percentage and deadline alone cannot establish a reliable plan.
Estimated lean mass is not a direct measurement of muscle. Bone and colleagues, 2017 demonstrate that water and glycogen changes can alter even DXA lean mass estimates.
