In Browser semantic wallpaper search over 16k+ wallpapers

Name: In Browser semantic wallpaper search over 16k+ wallpapers
Availability: InStock
Author: rdksu

by rdksu·Jun 16, 2026·3 points·0 comments

Visit Project View on HN

AI Analysis

●●●BangerWizardryDark Horse

16k image vector search running entirely in-browser with sqlite-vec and transformers.js.

Strengths

•CLIP text encoder inference happens client-side via transformers.js with no server.
•sqlite-vec extension enables vector search over 16k embeddings in browser SQLite.
•Zero backend infrastructure - all indexing and search happens on user's device.

Weaknesses

•Narrow use case limits broader applicability beyond wallpaper discovery.
•Initial load time for 16k embeddings could be substantial on slower devices.

Post Description

I indexed around 16k+ most-viewed wallpapers from wallhaven.cc using a CLIP-based model to generate image embeddings, and then hacked together a cool demo showing the vector search working client-side using sqlite-wasm compiled with the new and upcoming sqlite-vec extension, which allows me to do the vector search over the image index client-side! The inference over the actual text-encoder of CLIP also happens client-side using transformers.js :)