Google AI estimates body fat from a single selfie using a lightweight model on-device.

Google uses a single selfie to estimate body fat with a Pixel 7’s built-in model, processing the result in under two seconds without needing extra sensors. The lightweight MobileNetV3-small architecture—tuned to 1.2 million parameters and optimized for INT8 precision via TensorFlow Lite—runs in about 35 milliseconds on the Tensor G2 NPU, ensuring it doesn’t bottleneck the camera’s real-time pipeline.

The training pipeline begins with synthetic pre-training on SMPL-X 3D models, where lighting and backgrounds are randomized to simulate varied conditions. Afterward, the model fine-tunes on real DEXA scans, using mixup and cutout techniques to handle sparse labeled data. This dual-phase method helps refine predictions even when clinical-grade scans are limited.

Privacy is enforced by keeping inference strictly on-device, but a cleverly designed app with camera access could theoretically infer rough body-shape embeddings from repeated queries, raising concerns about membership attacks. Before returning the fat percentage, the model injects Gaussian noise (σ = 0.02), though this noise isn’t adjustable and lacks formal differential privacy guarantees.

DEXA scans remain the gold standard for distinguishing visceral from subcutaneous fat, but this method may still offer practical utility for general population screening or longitudinal self-monitoring. However, accuracy plummets for users with body fat percentages above 35% or below 18%, where clinical decisions demand precision.

For Android developers, Google provides a new BodyCompositionManager in Play Services 24.20, allowing integration via a single estimateBodyFat(bitmap) call. The response includes a BodyFatResult with a confidence interval, though iOS support remains exploratory. The model’s performance across diverse ethnicities and ages is unclear, as its initial DEXA dataset was mostly White (72%) and averaged 34 years old. Future studies may address potential biases, particularly for South Asian populations, where visceral fat risks persist at lower BMIs.

While this feature could enhance wellness dashboards with appropriate warnings, its role in insurance or clinical settings would require further validation and cross-platform verification before adoption.