Details
PinLanding uses a visual language model — an AI that can understand both images and text simultaneously — to analyze product images and generate descriptive attributes for each item (such as style, color, material, and occasion). A separate machine learning classifier then groups products with matching attributes into collections. Pinterest used GPT-4V (a vision-capable version of GPT-4) to generate high-quality attribute labels for training data, then assessed the quality of the resulting collections using a large language model as an automated judge. The system achieved a recall score of 99.7% in testing on a standard fashion product dataset, and improved average precision from 0.84 to 0.96 after refinements.
Have evidence about Pinterest's AI practices? Submit a report.
Report a Sighting →