3 comments

[ 3.6 ms ] story [ 17.8 ms ] thread
For years I've been taking screenshots of my phone as a way to record when a song is playing that I like.

I've also taken actual photos of "Now Playing" interfaces on other people's devices to try and save music that I've overheard playing (everything from car media interfaces to paper posters to pictures of Soundcloud pulled up on a laptop).

I've been doing so far without having a system to actually collect these photos into a playlist, but now that we have multi-modal models like GPT-4, I realized this is easy to build!

This script I created works surprisingly well despite only having 2 examples for GPT-4 to work off. I've fed it thousands of my screenshots from the last few years, and it's correctly identified tons of music. Next up we just need a script that can query TuneMyMusic or an equivalent Spotify/Apple Music playlist service to add all the detected songs to a playlist automatically.

The eventual dream is to be able to take a photo or screenshot of any kind of music interface in any app, and have it automatically end up in a playlist in my library within minutes (shazam-style, but for photos instead of audio).

Really neat idea, and with applications way beyond your original use case I suspect. :)
Yeah, even just being able to reliably detect the foreground app and all visible text opens up a ton of avenues for exploration.