Pinterest Aesthetic Analysis
I used image embeddings to look at the themes inside a Pinterest board, grouping similar images and keeping snapshots that could be compared over time.
Independent development. Python, CLIP embeddings and clustering.
I worked with a board as a collection of images, and the embedding step let the clustering operate on visual similarity while keeping the original images and metadata available for checking the groups, with the stored assignments making it possible to return from an abstract cluster to the images inside it.
The pipeline joined collection, embedding, clustering and comparison, with saved assignments and category refinements so a cluster could be inspected instead of accepting the generated label as an explanation by itself.
The checked-in August 2025 summary describes 171 images across three clusters, though that was a snapshot of one collection and not evidence that the same themes would remain stable on a different board.
The catalogue date follows the first preserved commit on 6 August 2025.
Outcome
An implemented image-analysis pipeline with saved cluster assignments and a dated comparison summary. The cover illustrates the grouping process without reproducing the collected photographs.