Depict: AI-powered search for Indian museum artworks using CLIP embeddings

1 dhrm1k 1 6/22/2025, 9:33:49 AM github.com ↗

Comments (1)

dhrm1k · 8h ago
Hi HN!

We built Depict (https://github.com/dhrm1k/depict) - an AI-powered search engine that lets you find Indian museum artworks using natural language descriptions like "painting of fruit" or "blue sculpture."

The project uses OpenAI's CLIP model to understand both text and images, creating embeddings that enable semantic visual search across collections from major Indian museums (National Gallery, National Museum Delhi, Indian Museum Kolkata, etc.).

Key technical details: - Scraped data from Museums of India government portal - Generated CLIP embeddings for artwork images - Built React frontend + Python backend on Modal AI - Similarity search using cosine distance in 512-dimensional space

The live demo currently includes ~2 museums due to compute costs, but the full pipeline is built to handle all 6 major collections. It was inspired by @ekzhang's dispict but focused specifically on making Indian cultural heritage more accessible.

Would love feedback on the search experience and any ideas for improving discoverability of museum collections!

Tech stack: Python, React, CLIP (ViT-B/32), Modal AI, h5py Repo: https://github.com/dhrm1k/depict