Open methodology

How we test AI virtual try-on

This page publishes the method before the result. The 20-garment benchmark has not been completed, so TryOn AI is not claiming an accuracy score. Publishing the protocol now makes it harder to change the rules after seeing the outputs.

Download the blank CSV scorecard

Current publication status

ItemStatusWhat that means
ProtocolPublishedThe method and blank scorecard are public on this page.
Approved reference photosPendingNo participant images will be published without permission.
20-garment runPendingNo benchmark result is claimed before all planned cases are complete.
Results and failed outputsPendingInputs and outputs will be released together.

The fixed test protocol

  1. 1

    Use two fixed base photos

    One torso photo is used for tops and jackets. One full-body photo is used for trousers and dresses. Pose, lighting and camera distance stay unchanged within each group.

  2. 2

    Test 20 garments

    The set contains five tops, five pairs of trousers, five dresses and five jackets. It includes plain colors, prints and details that are easy to inspect visually.

  3. 3

    Keep the source images

    Each record includes the garment source, the uncropped input and a note on image quality. A result without its input cannot be independently assessed.

  4. 4

    Score visible attributes

    Each output is checked for color, print, neckline, sleeve length and overall silhouette. The scorecard records yes, partly or no rather than inventing a confidence percentage.

  5. 5

    Publish failed cases

    Outputs with missing details, changed prints or implausible body geometry remain in the result set with a short failure note.

Interpretation rule

Visual preview is not fit prediction

The test measures visible preservation in generated images. It cannot establish the correct clothing size, comfort, fabric weight or real-world drape.

Until results exist

Inspect the product screens

The examples page shows what the current app interface supports and labels the limits of each example.

View the examples →