How Body Measurement AI Works — And Why It's Better Than Purchase History

Two approaches dominate the fashion tech conversation around fit intelligence: body measurement AI and purchase history modelling. Both aim to solve the same problem — helping customers find the right size — but they go about it in fundamentally different ways. Understanding the difference matters if you're trying to reduce returns and increase purchase confidence.

Purchase History Modelling: The Incumbent Approach

Purchase history modelling is the approach used by the largest online fashion platforms. The logic is straightforward: if a customer has bought a medium at your brand three times and kept all three items, recommend medium.

This works reasonably well under specific conditions. It requires a meaningful transaction history with a single brand. It works better for customers who buy frequently than for new customers. And it works better for brands with consistent sizing across their range than for brands with significant variation between product lines.

The limitations become clear quickly.

New customers have no history. Every customer starts with zero data. For brands with significant customer acquisition or high churn, a large proportion of their customer base at any given time has insufficient purchase history to generate a reliable recommendation.

Purchase history conflates fit with preference. If a customer kept a large because they prefer loose fits, that's a preference signal — not a fit signal. The model may recommend large for future purchases even when a medium would fit correctly.

Brand switching breaks the model entirely. A customer who is a confirmed medium at one brand may be a different size at yours if your size chart is cut differently. Purchase history from other brands is generally not shared and not comparable.

Body change isn't captured. A customer who has gained or lost weight since their last purchase will be recommended based on historical data that no longer reflects their current body.

Body Measurement AI: The Direct Approach

Body measurement AI takes a different approach entirely. Instead of inferring size from purchase behaviour, it reads the body directly.

The process uses computer vision — specifically pose estimation models that detect anatomical landmarks from a photograph. From a single full-body photo, the system identifies the positions of the shoulders, elbows, wrists, hips, knees, and ankles. From those landmark positions, combined with an estimated height derived from the proportions of the detected skeleton, it calculates real measurements.

What gets measured:

• Shoulder width — distance between the left and right shoulder landmarks, scaled to centimetres.
• Chest circumference — derived from shoulder width and calibrated against population data.
• Torso length — vertical distance from shoulder midpoint to hip midpoint.
• Waist circumference — derived from chest and torso proportions.
• Hip width — distance between left and right hip landmarks, scaled and circumference-converted.
• Arm length — cumulative distance from shoulder to elbow to wrist.
• Inseam — distance from hip landmark to ankle.
• Neck circumference — derived from chest measurement.

These eight measurements are then mapped to the brand's specific size chart — not a generic standard — to generate a size recommendation.

GarmCheck

UI screenshotGarmCheck measurement panel — 8 measurements with "Detection complete"

Why This Is Fundamentally Better

The difference between purchase history and body measurement isn't just technical — it's conceptual.

Purchase history answers the question: "What size has this customer bought before?" Body measurement answers the question: "What size will fit this customer's actual body today?"

These are different questions. For many customers, the answers are the same. For a significant proportion — new customers, customers who have changed shape, customers whose preference doesn't match their ideal fit — the answers diverge. And it's in that divergence that most fit-related returns originate.

Body measurement is also brand-agnostic. A customer's shoulder width is a physical fact that doesn't change based on which brand they've previously purchased from. This makes body measurement particularly valuable for customer acquisition — it works just as well for a first-time buyer as for a repeat customer.

The Visual Layer: Why Measurement Alone Isn't Enough

One thing purchase history modelling and basic size recommendation tools share is that they give customers a number — a recommended size — without giving them visual confidence.

The measurement-based size recommendation is more accurate than a chart lookup. But it still asks the customer to take a leap of faith: "the AI says medium, I'll trust it."

The most effective implementation combines body measurement with garment rendering: using the detected body landmarks as the spatial reference for compositing the garment onto the customer's photo. The customer doesn't just receive a size recommendation — they see the garment on their body, with the fabric draping realistically, the shoulders aligned to their shoulder width, and the hem sitting at the correct position relative to their actual proportions.

GarmCheck

UI screenshotFull GarmCheck result — try-on image left, size recommendation with M highlighted right

This visual confirmation is what moves the needle on purchase confidence. The combination of accurate measurement and photorealistic rendering removes uncertainty at the two levels where it exists: "will it fit?" and "will it look right?"

The Accuracy Question

No computer vision system achieves perfect measurement accuracy from a single photo. Variables like camera angle, photo crop, clothing worn in the image, and lighting all affect the precision of landmark detection.

The relevant question isn't whether the measurements are perfect. It's whether they're accurate enough to reliably place customers in the correct size bucket — and consistent enough that the same customer gets the same recommendation on repeat visits.

In practice, body measurement AI achieves sufficient accuracy to match the right size the majority of the time. The remaining error is typically boundary cases — customers who sit between sizes, where the recommendation of either adjacent size would be defensible. These cases are handled with advisory overlays and fit type logic rather than hard blocking.

The comparison point isn't perfection. It's the alternative: customers self-measuring inaccurately, or relying on purchase history that may not reflect their current body or the current brand's sizing. Against that baseline, body measurement AI is meaningfully more accurate — and meaningfully better at reducing fit-related returns.