Google Research just published a system that estimates how much body fat you're carrying, and where it's distributed, from two ordinary photos taken with a phone camera — and in Google's own testing, it lands closer to a clinical DXA scan than the bioelectrical sensor built into your smartwatch does.
How two photos beat a $300 wearable sensor
The mechanics are simpler than the accuracy numbers suggest. You stand for two ordinary 2D photos — one facing the camera, one from the side. A ResNet-50 image model, the same general architecture used in a lot of computer vision work and initialized with standard ImageNet weights, reads your body shape from those two frames. Google then fuses that visual read with basic inputs anyone already knows — sex, height, weight, BMI — through a dense neural layer to produce the final estimate. No sensor contact, no scale, no bioelectrical current run through your body.
Stand for two photos: one facing the camera, one from the side.
2
A ResNet-50 image model reads body shape and composition from the two frames.
3
The model fuses that read with your sex, height, weight and BMI.
4
Output: body fat percentage, fat distribution ratios, and an insulin-resistance risk score.
How PhotoScan stacks up
PhotoScan (2 photos)
Method
2.13-2.15 percentage points MAE
Body-fat error vs. DXA
A phone camera, no extra hardware
Smartwatch BIA sensor
Method
Higher error than PhotoScan in Google's testing
Body-fat error vs. DXA
A $250-400 wearable
Clinical DXA scan
Method
Reference standard (what PhotoScan is measured against)
Body-fat error vs. DXA
A hospital or clinic visit
Method
Body-fat error vs. DXA
What it needs
PhotoScan (2 photos)
2.13-2.15 percentage points MAE
A phone camera, no extra hardware
Smartwatch BIA sensor
Higher error than PhotoScan in Google's testing
A $250-400 wearable
Clinical DXA scan
Reference standard (what PhotoScan is measured against)
A hospital or clinic visit
The catch: it's a research prototype, not a Pixel feature yet
Google Research's own authors, Cassie Zhou and Ahmed Metwally, describe PhotoScan as an "investigational deep learning framework" — the paper is a proof of concept, not a product announcement. There's no confirmed timeline for it landing in Google Fit, on the Pixel Watch, or as a Pixel camera feature. That distinction matters more than the headlines chasing it suggest: the accuracy numbers are genuinely strong, but strong research results and a shipped, privacy-reviewed consumer feature are two different bars, and Google has quietly shelved health research before it reached either.
Body composition tracking has mostly lived on wearables and smart scales — PhotoScan tries to do it with just a camera. · Unsplash
Part of a much bigger pattern
PhotoScan lands right in the middle of Google trying to make AI the actual differentiator on Pixel hardware rather than a marketing line — see the storage and pricing shakeup around the Pixel 11 launch, or the broader question of how much AI your phone can actually run without offloading to the cloud. Apple's answer to the same pressure was paying Google a reported $1 billion to bring Gemini into Siri rather than building health-camera research of its own. Whoever gets a feature like this into a shipping product first, cleanly and with a privacy story that holds up, has a real hook — this isn't a chatbot gimmick, it's a reason to actually reach for your phone.
Is PhotoScan available on Pixel phones right now?
No — it's a research prototype described in a Google Research paper, not a shipping feature in Google Fit, Pixel Watch, or any Pixel phone yet.
How accurate is it compared to a DXA scan?
In Google's validation testing, PhotoScan's body-fat-percentage estimate landed within about 2.1-2.15 percentage points of a clinical DXA scan on average, and it predicted insulin resistance risk with 76% accuracy — close to DXA-based estimates.
What photos does it actually need?
Two ordinary 2D smartphone photos — one from the front, one from the side. No special lighting rig or sensor.
Is this different from the Pixel Watch's existing health features?
Yes. The Pixel Watch already tracks metrics like Insulin Resistance Trends using its own sensors over time; PhotoScan is a separate, camera-only approach that doesn't require wearing anything.
Watch for this to resurface at a Pixel hardware event or in a Google Fit changelog rather than as a standalone launch — that's the usual path Google Research work takes when it graduates into a real feature. Until then, treat the accuracy numbers as a real signal of what's coming, not a feature you can use today.
