Pinterest Launches AI Feature to Find and Refine Ideas Using Images

Pinterest has announced a series of new AI-powered visual search features that will make it easier for users to find exactly what they're looking for through images. The company says that keyword searches can sometimes be difficult to interpret when users are trying to identify a specific style or aesthetic, and that focusing on visual search is a natural fit for the visual platform.

Here are the key features worth noting:

Detailed image analysis:

When viewing a Pin, Pinterest's AI will suggest keywords that describe parts of the image like "overall aesthetic," "color palette," "style," or "product category." This helps users understand their interests and discover related ideas/products. A moving light effect is also added to highlight interactive elements.

1. Fine-tuning:

Pinterest added a toolbar that allows users to filter search results by specific criteria (color, style, material.) so that users can quickly find the most satisfactory version.

2. Expand the "image search" function:

Users can simply tap and hold any Pin on the main feed to activate image search, instead of having to switch to a separate tab.

According to Pinterest, these features are built on the following advanced models:

1. Visual Language Models (VLMs): A form of generative AI that expands the 'visual vocabulary' to describe stylistic ideas.
2. Multimodal Embedding Model: Combines images and text to find content that precisely matches the user's aesthetic taste.

Additionally, to address concerns about 'AI slop,' Pinterest is starting to label Pins that are created or significantly altered with AI with an 'AI Edited' label, to increase transparency.

The new features will roll out first to women's fashion content in the UK, US, and Canada, then expand to more categories and countries.

With this move, Pinterest continues to affirm its position as a platform connecting creative ideas through AI technology, while solving the "hard to put into words" search problem of users.

Samuel Daniel

Update 26 May 2025