Can ChatGPT accurately estimate Body Fat Percentage (BFP) from photos? Surprisingly, yes.
How close can AI get to a clinical scan just from a photo? This close.
The most advanced AI model today might be able to tell you how fat you are, accurately. And this is important.
From my first crude test, GPT-4o's ability to estimate body fat percentage from photos rivals even gold-standard tools like DEXA scans.
Why this matters: As a direct measurement of adiposity (and inversely, lean body mass), BFP is a more precise and meaningful metric for health than Body Mass Index (BMI), but it’s typically inaccessible without specialized tools or less affordable.
From my training as a strength coach, I had learned that a trained eye (with expert intuition) can judge BFP roughly as accurately as other methods. I figured this might equally apply to ChatGPT with its ability to analyze images.
To test this, I fed the images from Menno Henselmans' "Visual Guides to Body Fat Percentage" (male and female) into ChatGPT and compared its estimates to the original numbers.
The results are in:
Men: median signed error: +0.8%, median absolute error: 2.4%
Women: median signed error: +3.5%, median absolute error: 5.7% (some underestimation in the higher percentages for women, but BFP is in general harder to judge for females)
With expensive methods like DEXA scans having an error around 2%, this is pretty good. And GPT-4o always correctly placed each case into the right rough category, from “bodybuilding stage-ready” to “obese II.”
While not a medical diagnosis (!), a photo and ChatGPT might give you a decent starting point for understanding and improving your health. BMI is outdated. DEXA is expensive. ChatGPT might be the new middle ground.
This is very interesting!
I wonder whether these photos were used for training the model. If that happened and the corresponding BFP sits close to the pictures, the model would probably make the association. Well, if we ignore the privacy questions, one probably could run the same tests with its own photos.
An interesting post! Have you performed any sort of k-fold cross validation?
I've used a 5-fold validation before. Shuffle your images randomly, and then break your data set into 5 subsets. Repeat the training and evaluation steps of your experiment 5 times, each time leaving out a different subset from the training corpus, and then evaluate the trained model against the images that were in the left out set.
Eg, if you shuffle your images and break them out into subsets A through E:
Round 1: Train on A, B, C, D; evaluate model against E.
Round 2: Train on A, B, C, E; evaluate model against D.
Round 3: Train on A, B, D, E; evaluate model against C.
Round 4: Train on A, C, D, E; evaluate model against B.
Round 5: Train on B, C, D, E; evaluate model against A.
From there you can average the results from all 5 rounds to get a better idea of how well your model performs against images it's never seen before.