README.md

---
title: Wonder Finder
emoji: 🌍
colorFrom: yellow
colorTo: yellow
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: mit
---

# 🌍 Wonder Finder

Visual recommender for the 12 Wonders of the World, powered by CLIP embeddings.

## What it does
- **Image search:** upload a travel photo → get the 3 most visually similar wonders
- **Text search:** describe a place in natural language → get the 3 closest matching wonders

## How it works
1. The catalog (11,544 images across 12 wonder classes) is pre-embedded using CLIP ViT-B/32.
2. User input (image or text) is embedded into the same 512-D space.
3. Cosine similarity ranks the catalog; top 3 results are returned with a diversity filter to avoid duplicates.

## Dataset
[chavajaz/wonders_dataset](https://huggingface.co/datasets/chavajaz/wonders_dataset) — CC0-1.0 licensed, ~960 images per class on average.

## Model
[openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) — chosen for its joint image-text embedding space, which enables both image and text input through a single model.

## Cluster analysis
K-Means at k=12 on the embeddings achieved **ARI = 0.890** and **NMI = 0.927** against ground-truth wonder labels, indicating CLIP's pretrained space already separates the 12 wonders almost perfectly without supervision.

Built as the final project for [Course Name] Assignment 3.