Find Art by Similarity

For this project I built a visual similarity search across the entire portfolio. Customers upload any image — a piece they love, a photo from a magazine, a screenshot — and the system returns the artworks that are closest in visual character.

At the core is a vision encoder from a contrastively trained Vision-Language model — the kind of training that produces embeddings with a deep understanding of visual style, not just content. At query time, only the image side of the model is used: the uploaded image is encoded into a vector, and a nearest-neighbour search across the precomputed portfolio embeddings returns the closest matches.

The result is a discovery experience that works the way intuition does — not through tags or categories, but through visual affinity.